{"id":4209,"date":"2026-09-21T14:40:27","date_gmt":"2026-09-21T14:40:27","guid":{"rendered":"https:\/\/projectfifty4.com\/ai-search-visibility-energy\/"},"modified":"2026-09-21T14:45:32","modified_gmt":"2026-09-21T14:45:32","slug":"ai-search-visibility-energy","status":"publish","type":"post","link":"https:\/\/projectfifty4.com\/fr\/ai-search-visibility-energy\/","title":{"rendered":"AI Search Visibility for Energy Companies"},"content":{"rendered":"<p><strong>Buyers now ask ChatGPT, Perplexity, Gemini and Claude who to hire and what to buy. Those assistants cite a different set of sources than Google does, so ranking well on Google does not put you in the answer. This is what actually moves AI visibility, what does not, and how Project 54 measures it for energy companies.<\/strong><\/p>\n<h2>What is AI search visibility, and how do energy companies improve it?<\/h2>\n<p>AI search visibility is whether an AI assistant names and cites your company when a buyer asks it a question. It is measured separately from Google rankings because the two barely overlap: peer-reviewed measurement of 1,000 queries across five engines found the share of an assistant&#8217;s cited domains that also appear in Google&#8217;s top 10 was 4.0 percent for GPT-4o, 11.1 percent for Gemini, 12.6 percent for Claude and 15.2 percent for Perplexity. Improving it rests on three things. First, being fetchable: the retrieval crawlers that power AI answers do not execute JavaScript, and blocking the wrong bot removes you from a product&#8217;s answers entirely. Second, being mentioned elsewhere: across 75,000 brands, branded web mentions correlated with AI Overview presence at 0.664 against 0.218 for backlinks, and research published at EMNLP 2025 found model citation preference tracks the source&#8217;s identity rather than the article&#8217;s content. Third, being specific: on narrow, technical questions where no encyclopaedia entry exists, a specialist source can win, which is where energy companies have an advantage they rarely use.<\/p>\n<h2>Points cl\u00e9s \u00e0 retenir<\/h2>\n<ul>\n<li>Google rank does not buy AI citation. Measured overlap between an assistant&#8217;s citations and Google&#8217;s top 10 runs from 4 percent on GPT-4o to 15 percent on Perplexity. Google&#8217;s own AI Overviews are the exception because they run on Google&#8217;s index, and even there only 37.1 percent of citations came from the top 10 in a March 2026 measurement of 863,000 result pages.<\/li>\n<li>The most common self-inflicted injury is blocking the wrong crawler. GPTBot collects training data. OAI-SearchBot is what makes you eligible to appear in ChatGPT search. Blocking GPTBot changes nothing about your visibility in ChatGPT. Blocking OAI-SearchBot removes you from it.<\/li>\n<li>Retrieval crawlers do not run JavaScript. Log analysis of GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot and Meta&#8217;s agent found none of them execute it. Content that only exists after the page renders is invisible to them.<\/li>\n<li>Off-site mentions matter more than on-site markup. Schema markup showed no citation uplift in a controlled study of 1,885 pages against roughly 4,000 matched controls, while branded web mentions were the strongest measured correlate of AI Overview presence across 75,000 brands.<\/li>\n<li>Two on-page tactics have measured causal support: adding attributed quotations and adding named statistics, at plus 43 percent and plus 33 percent on the research benchmark. The same research found the effect reverses for sources already ranked first, which means these tactics pay most for challengers rather than incumbents.<\/li>\n<li>Grounded and ungrounded results are not the same thing and should never be reported as one number. An API model answering from memory is not the product your buyer uses. Any claim about what happens when someone asks an assistant has to rest on a run with live retrieval enabled.<\/li>\n<\/ul>\n<h2>The shortlist is being drafted by software<\/h2>\n<p>For a decade the change in B2B buying was self-service: buyers researched independently and contacted sales late. The change now is delegation. Buyers ask an assistant, and the assistant reads, compares and shortlists on their behalf.<\/p>\n<p>The mechanics differ by engine and it matters. <a href=\"https:\/\/developers.google.com\/search\/docs\/fundamentals\/ai-optimization-guide\" target=\"_blank\" rel=\"noopener nofollow\">Google states that AI Overviews and AI Mode rely on its core Search ranking systems to retrieve pages from its Search index<\/a>, combined with a query fan-out that generates several related searches at once. <a href=\"https:\/\/www.perplexity.ai\/hub\/blog\/architecting-and-evaluating-an-ai-first-search-api\" target=\"_blank\" rel=\"noopener nofollow\">Perplexity has published that it runs its own index of more than 200 billion URLs<\/a>, decomposes documents into self-contained spans that are retrieved and ranked individually, and applies cross-encoder rerankers to the final selection. OpenAI says ChatGPT search draws on third-party search providers alongside its own indexed and cached web content.<\/p>\n<p>Three consequences follow for an energy supplier. The engines are reading different indexes, so visibility is not one number. They retrieve passages rather than pages, so a page that is strong overall but vague in its middle sections gets passed over. And the citation you see is a filtered subset of what was consulted: OpenAI&#8217;s own API documentation notes that the list of sources the model consulted is often larger than the number of citations shown.<\/p>\n<p>There is also a quality problem nobody selling this should hide. The standard academic audit of citation accuracy, <a href=\"https:\/\/arxiv.org\/abs\/2304.09848\" target=\"_blank\" rel=\"noopener nofollow\">Liu, Zhang and Liang, published at EMNLP Findings in 2023<\/a>, examined four generative search engines and found only 51.5 percent of generated sentences were fully supported by their cited sources, and 74.5 percent of citations supported the statement they were attached to. Being cited is not the same as being described correctly, which is why we measure both.<\/p>\n<h2>Training bots and retrieval bots are not the same thing<\/h2>\n<p>This is the cheapest and most commonly mishandled part of the whole discipline. Many firms blocked AI crawlers wholesale on legal advice between 2023 and 2025. Some of them removed themselves from AI answers without intending to, and have no idea.<\/p>\n<p>The distinction is between crawlers that gather training data and crawlers that build the retrieval index an assistant reads at answer time. <a href=\"https:\/\/developers.openai.com\/api\/docs\/bots\" target=\"_blank\" rel=\"noopener nofollow\">OpenAI documents four separate agents<\/a>: GPTBot for model training, OAI-SearchBot for ChatGPT search eligibility, ChatGPT-User for user-initiated fetches, and OAI-AdsBot. Blocking GPTBot excludes you from future training sets and does nothing to your presence in ChatGPT search. Blocking OAI-SearchBot removes you from ChatGPT search answers.<\/p>\n<p><a href=\"https:\/\/support.claude.com\/en\/articles\/8896518\" target=\"_blank\" rel=\"noopener nofollow\">Anthropic documents three<\/a>: ClaudeBot for training, Claude-User for user-initiated retrieval, and Claude-SearchBot for search indexing. All three obey robots.txt. <a href=\"https:\/\/docs.perplexity.ai\/docs\/resources\/perplexity-crawlers\" target=\"_blank\" rel=\"noopener nofollow\">PerplexityBot builds the search index and explicitly does not train models<\/a>, so blocking it removes you from Perplexity results. Google&#8217;s Google-Extended is a control token rather than a separate crawler, and its documentation now covers grounding in Gemini as well as training, so blocking it can remove you from Gemini answers while leaving Search untouched.<\/p>\n<p>Two further traps. Verify crawlers by the published IP ranges rather than the user-agent string, which is trivially spoofed. And check your content delivery network separately from your robots file: from 15 September 2026 Cloudflare changed its defaults for new customers, new sites of existing customers and existing free-plan customers to block mixed-use AI crawlers on ad-bearing pages. A firm can have a permissive robots.txt and still be invisible because something upstream is returning a 403.<\/p>\n<p>The rendering point closes this section. Log analysis published by Vercel and MERJ found that none of GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot or Meta&#8217;s agent execute JavaScript. GPTBot fetched JavaScript files in around 11.5 percent of requests and ClaudeBot in around 23.8 percent, but read them as text rather than running them. If your capability statement, specification table or price is built client-side, it does not exist as far as those crawlers are concerned. The test is the page source, not the browser inspector.<\/p>\n<h2>Mentions first, structure second, markup last<\/h2>\n<p>The strongest measured signal is not on your website. Across 75,000 brands, <a href=\"https:\/\/ahrefs.com\/blog\/ai-overview-brand-correlation\/\" target=\"_blank\" rel=\"noopener nofollow\">Ahrefs found branded web mentions correlated with AI Overview presence at 0.664<\/a>, against 0.527 for branded anchors, 0.392 for branded search volume, 0.326 for domain rating and 0.218 for backlinks. Brands in the top quartile by web mentions averaged 169 AI Overview mentions; the next quartile down averaged 14. Ahrefs states plainly that this is correlation rather than proof of causation, and we repeat that caveat rather than burying it.<\/p>\n<p>It is reinforced by a mechanism rather than only a correlation. Research presented at <a href=\"https:\/\/aclanthology.org\/2025.emnlp-main.872\/\" target=\"_blank\" rel=\"noopener nofollow\">EMNLP 2025 found that language model citation preference tracks the identity of the outlet rather than the content of the article<\/a>. The models could barely infer an article&#8217;s stance from its text but recognised the source almost perfectly. If that holds, then who you are is doing more work than what any individual page says, which is an uncomfortable finding for anyone selling page-level optimisation.<\/p>\n<p>On the page itself, two tactics have measured causal support. The benchmark study of generative engine optimisation, <a href=\"https:\/\/arxiv.org\/abs\/2311.09735\" target=\"_blank\" rel=\"noopener nofollow\">published at KDD 2024 across 10,000 queries<\/a>, measured adding attributed quotations at plus 43 percent and adding named statistics at plus 33 percent on its visibility metric. Keyword stuffing scored below the baseline. Two caveats that are almost never quoted: the engine was simulated rather than live, and the effect reverses by rank. For a source already ranked first, quotations cost 22.9 percent and statistics cost 20.6 percent. For a source ranked fifth they gain 99.7 percent and 97.9 percent. These are challenger tactics, and most energy suppliers are challengers.<\/p>\n<p>Freshness helps, unevenly. Across 16.975 million cited URLs, AI-cited content was 25.7 percent fresher by publication date, but the effect concentrated in ChatGPT and was essentially absent in Google AI Overviews, which cited content 16 days older than the organic results. Peer-reviewed measurement found the median age of cited articles varies hugely by category, from 62 days for Claude in consumer electronics to 493 days for Google in automotive. Recency is a lever in some categories and noise in others.<\/p>\n<p><strong>Be fetchable<\/strong> Correct robots rules by crawler, no content delivery network blocking retrieval bots, and substantive content in the server-rendered HTML rather than built by JavaScript after load.<\/p>\n<p><strong>Be mentioned<\/strong> Independent pages that name your firm are the strongest measured correlate. Trade press, association directories, conference listings, industry roundups. This is earned, not marked up.<\/p>\n<p><strong>Be specific<\/strong> You will not out-cite Wikipedia on broad questions. You can win narrow technical questions where no encyclopaedia entry exists, which in energy is most of the questions that matter commercially.<\/p>\n<p><strong>Be checkable<\/strong> Named figures with a source and a date, attributed quotes, explicit methodology. The two tactics with measured causal support are both forms of showing your working.<\/p>\n<h2>The part of this market that is selling folklore<\/h2>\n<p>We publish this section because a buyer cannot evaluate a supplier in this field without it, and because most of what is sold under the heading of AI optimisation has no evidence behind it.<\/p>\n<p>Schema markup does not buy citation. A difference-in-differences study tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 against roughly 4,000 matched controls. AI Mode citations moved 2.4 percent and ChatGPT 2.2 percent, both inside noise, and AI Overview citations fell 4.6 percent. <a href=\"https:\/\/developers.google.com\/search\/docs\/fundamentals\/ai-optimization-guide\" target=\"_blank\" rel=\"noopener nofollow\">Google&#8217;s own guidance says structured data is not required and there is no special markup you need to add<\/a>. Schema still earns its place for entity disambiguation and Google rich results. It is not an AI visibility lever and should not be sold as one.<\/p>\n<p>FAQPage markup in particular is now close to worthless. Google removed FAQ rich results from Search entirely on 7 May 2026. The text of a well-written question and answer still helps, because it is visible content that retrieves cleanly. The markup wrapper no longer does anything.<\/p>\n<p>The file called llms.txt is a bet rather than an optimisation. Google&#8217;s John Mueller has said no AI system currently uses it, and Gary Illyes has said Google does not support it and has no plans to. No major AI company has committed to reading it. A survey of 137,210 domains in May 2026 found 28 percent published one and 97 percent of those received zero requests for it. It costs almost nothing to publish and we do not object to it, but anyone presenting it as a core deliverable is selling you a file nothing reads.<\/p>\n<p>Three more to retire. Rewriting content in an artificially AI-friendly voice: Google states you do not need to write in a specific way for generative search, and the benchmark study&#8217;s own fluency optimisation sat mid-table and reversed for high-ranked sources. Breaking content into small chunks for AI: Perplexity and Bing chunk internally regardless of what you do, and Google says there is no requirement to do it and no ideal page length. And the widely circulated claim that 87 percent of ChatGPT citations rank in Bing&#8217;s top 10: the original study covered 100 queries against the top 20 on a superseded product, and larger measurements of 15,000 and 118,931 queries put the figure at 8.1 and 6.82 percent.<\/p>\n<h2>Narrow questions, gated proof and a procurement layer<\/h2>\n<p>Three features of energy change the problem. The first is question shape. General-interest queries are dominated by encyclopaedic and community sources: Wikipedia accounted for 29.7 percent of ChatGPT&#8217;s top-cited pages in one study and Reddit for 16.8 percent of its citation share. A specialist supplier will not beat those on broad questions. But the questions that decide energy purchases are not broad. What certification does this duty require, which vendors are registered with this operator, how is an in-country value score calculated, what did this award actually include. There is no encyclopaedia entry for those, and whoever answers them properly is the available source.<\/p>\n<p>The second is that the proof lives behind gates. Energy suppliers put their most citable material, the specifications, the certifications, the reference installations, the local content position, into PDFs, portals and sales conversations. Retrieval crawlers read HTML well, PDFs poorly, and gated content not at all. A supplier can be the most qualified firm for a duty and be entirely absent from the machine-readable record of that duty.<\/p>\n<p>The third is the procurement layer. In energy the buyer often is not the payer. The operator writes the specification and maintains the approved vendor list, the contractor places the order, and a local content scheme such as Saudi Aramco&#8217;s iktva or ADNOC&#8217;s In-Country Value programme scores the bid. An assistant asked to shortlist suppliers is reasoning across all of that, and the firms that publish their registration status, certification scope and local content position as plain text are the ones it can actually place. We wrote about the underlying structure in <a href=\"\/fr\/channel-partner-marketing-energy-b2b\/\">channel and partner marketing in energy B2B<\/a>.<\/p>\n<p>There is a forward-looking reason to treat this as urgent rather than interesting. Gartner forecasts that by 2028 around 90 percent of B2B buying will be intermediated by AI agents, with more than 15 trillion dollars flowing through agent exchanges, while separately forecasting that by 2030, 75 percent of B2B buyers will prefer sales experiences that prioritise human interaction over AI. Both can be true. The qualification and shortlisting layer becomes machine-mediated while negotiation stays human, which means the machine-readable layer decides who gets into the conversation at all.<\/p>\n<h2>Grounded and ungrounded are two different numbers<\/h2>\n<p>The single most important rule in this discipline, and the one most often broken, is that an API model answering with no web access is not the product your buyer uses. A model can name your firm from training memory while the live consumer product, running retrieval, never surfaces you. Reporting those together produces a flattering number that describes nothing.<\/p>\n<p>Project 54 reports two separate figures. Retrieval visibility is measured with live web access enabled, which is what actually happens when a buyer asks. Model memory is measured without it, which tells you what the model has absorbed about you historically. They move for different reasons and they need different remedies: retrieval visibility responds to being fetchable and being mentioned, model memory responds only slowly and only to sustained presence.<\/p>\n<p>The method is a fixed prompt panel rather than a one-off check, because run-to-run variance is large and any single result is close to meaningless. We fix a set of prompts a real buyer would type, in each language the market actually uses, run them on a fixed schedule from a clean logged-out session with location held constant, and record per run whether the firm was named, whether it was cited with a link, where in the answer it appeared, which competitors were named, and which URLs were cited. Trend is the finding. A single run is an anecdote.<\/p>\n<p>One discipline check that costs nothing and catches a lot: cross-reference the URLs an engine cited against your own server logs. If a page is being cited but was never fetched by that engine&#8217;s crawler, the citation came from training memory rather than retrieval, and the work required to defend it is completely different from the work required to earn it.<\/p>\n<p>This is what the <a href=\"\/fr\/audit-de-visibilite\/\">Audit de la capacit\u00e9 de recherche \u00e9nerg\u00e9tique<\/a> does as a paid engagement, and the measurement discipline above is the product. The reason we publish the method rather than hiding it is that in a field this full of unverifiable claims, the method is the differentiator.<\/p>\n<h2>FAQ<\/h2>\n<h3>Is AI search visibility the same as SEO?<\/h3>\n<p>No, and treating them as one is the most common mistake. They overlap for Google AI Overviews, which run on Google&#8217;s own index, but barely overlap for the assistants. A peer-reviewed study of 1,000 queries across five engines measured the share of an assistant&#8217;s cited domains that also appear in Google&#8217;s top 10 at 4.0 percent for GPT-4o, 11.1 percent for Gemini, 12.6 percent for Claude and 15.2 percent for Perplexity. Good SEO helps you on Google&#8217;s AI surfaces and does relatively little for the rest. The disciplines share some foundations, mainly crawlability and clear content, and diverge sharply after that.<\/p>\n<h3>What is the difference between GEO, AEO and AI search visibility?<\/h3>\n<p>They are largely the same thing under different labels. Generative engine optimisation, or GEO, comes from a 2024 academic paper and refers to increasing a source&#8217;s prominence in generated answers. Answer engine optimisation, or AEO, is the more common commercial term for the same goal. AI search visibility is the plainest description and the one we prefer, because it names the outcome rather than a technique. Be careful with any supplier who presents these as three separate services with three separate fees.<\/p>\n<h3>Does blocking AI crawlers protect our content?<\/h3>\n<p>It depends entirely on which crawler, and the distinction is where most damage is done. Blocking a training crawler such as GPTBot or ClaudeBot keeps your content out of future model training and has no effect on whether you appear in that company&#8217;s AI search product. Blocking a retrieval crawler such as OAI-SearchBot, Claude-SearchBot or PerplexityBot removes you from those products&#8217; answers. Firms that blocked AI bots wholesale on legal advice frequently blocked both without meaning to. It is free to check and takes minutes.<\/p>\n<h3>Do we need llms.txt?<\/h3>\n<p>There is currently no credible evidence any production AI system reads it. Google&#8217;s John Mueller has stated that no AI system uses it, and a survey of 137,210 domains in May 2026 found that 97 percent of published llms.txt files received zero requests. It is cheap and harmless to publish, so we do not argue against it, but it should never displace work on crawler access, server-side rendering or off-site presence. If a supplier&#8217;s AI visibility proposal leads with llms.txt, ask what evidence they have that anything reads it.<\/p>\n<h3>How long does it take to improve AI visibility?<\/h3>\n<p>The technical layer is fast and the entity layer is slow. Crawler access, rendering and content structure are fixable in weeks and their effect appears at the next crawl. Being mentioned across independent sources is a programme rather than a project, and the evidence suggests it is the layer that matters most. Model memory, as distinct from retrieval, lags further still because it depends on training cycles. Anyone promising to move your AI visibility in a fixed short window is either talking only about the technical layer or is not being straight with you.<\/p>\n<h3>What does Project 54&#x27;s Energy Findability Audit cover?<\/h3>\n<p>It measures AI visibility across a fixed prompt panel, in multiple languages, against named competitors, reporting retrieval visibility and model memory as two separate numbers rather than one blended figure. It also covers the technical layer that decides whether you can be retrieved at all: crawler access by agent, rendering, and whether your capability content exists in machine-readable form. It is a fixed-price engagement from 3,500 euros excluding value added tax, with the level set by the number of prompts, engines, competitors and languages covered. Full scope and pricing are on the audit page.<\/p>","protected":false},"excerpt":{"rendered":"<p>Buyers now ask ChatGPT, Perplexity, Gemini and Claude who to hire and what to buy. Those assistants cite a different set of sources than Google does, so ranking well on Google does not put you in the answer. This is what actually moves AI visibility, what does not, and how Project 54 measures it for<\/p>","protected":false},"author":12,"featured_media":4205,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"p54_article_data":"{\"meta\": {\"kicker\": \"Service \u00b7 AI Search Visibility\", \"topics\": [\"Marketing\", \"AI Visibility\"], \"title\": \"AI Search Visibility for Energy Companies\", \"dek\": \"Buyers now ask ChatGPT, Perplexity, Gemini and Claude who to hire and what to buy. Those assistants cite a different set of sources than Google does, so ranking well on Google does not put you in the answer. This is what actually moves AI visibility, what does not, and how Project 54 measures it for energy companies.\", \"date\": \"2026-09-21\", \"readTime\": \"13 min read\", \"author\": \"Project 54\"}, \"quickAnswer\": {\"q\": \"What is AI search visibility, and how do energy companies improve it?\", \"a\": \"AI search visibility is whether an AI assistant names and cites your company when a buyer asks it a question. It is measured separately from Google rankings because the two barely overlap: peer-reviewed measurement of 1,000 queries across five engines found the share of an assistant's cited domains that also appear in Google's top 10 was 4.0 percent for GPT-4o, 11.1 percent for Gemini, 12.6 percent for Claude and 15.2 percent for Perplexity. Improving it rests on three things. First, being fetchable: the retrieval crawlers that power AI answers do not execute JavaScript, and blocking the wrong bot removes you from a product's answers entirely. Second, being mentioned elsewhere: across 75,000 brands, branded web mentions correlated with AI Overview presence at 0.664 against 0.218 for backlinks, and research published at EMNLP 2025 found model citation preference tracks the source's identity rather than the article's content. Third, being specific: on narrow, technical questions where no encyclopaedia entry exists, a specialist source can win, which is where energy companies have an advantage they rarely use.\"}, \"takeaways\": [\"Google rank does not buy AI citation. Measured overlap between an assistant's citations and Google's top 10 runs from 4 percent on GPT-4o to 15 percent on Perplexity. Google's own AI Overviews are the exception because they run on Google's index, and even there only 37.1 percent of citations came from the top 10 in a March 2026 measurement of 863,000 result pages.\", \"The most common self-inflicted injury is blocking the wrong crawler. GPTBot collects training data. OAI-SearchBot is what makes you eligible to appear in ChatGPT search. Blocking GPTBot changes nothing about your visibility in ChatGPT. Blocking OAI-SearchBot removes you from it.\", \"Retrieval crawlers do not run JavaScript. Log analysis of GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot and Meta's agent found none of them execute it. Content that only exists after the page renders is invisible to them.\", \"Off-site mentions matter more than on-site markup. Schema markup showed no citation uplift in a controlled study of 1,885 pages against roughly 4,000 matched controls, while branded web mentions were the strongest measured correlate of AI Overview presence across 75,000 brands.\", \"Two on-page tactics have measured causal support: adding attributed quotations and adding named statistics, at plus 43 percent and plus 33 percent on the research benchmark. The same research found the effect reverses for sources already ranked first, which means these tactics pay most for challengers rather than incumbents.\", \"Grounded and ungrounded results are not the same thing and should never be reported as one number. An API model answering from memory is not the product your buyer uses. Any claim about what happens when someone asks an assistant has to rest on a run with live retrieval enabled.\"], \"sections\": [{\"id\": \"what-changed\", \"q\": \"What has actually changed in how energy buyers find suppliers?\", \"h\": \"The shortlist is being drafted by software\", \"p\": [\"For a decade the change in B2B buying was self-service: buyers researched independently and contacted sales late. The change now is delegation. Buyers ask an assistant, and the assistant reads, compares and shortlists on their behalf.\", \"The mechanics differ by engine and it matters. <a href=\\\"https:\/\/developers.google.com\/search\/docs\/fundamentals\/ai-optimization-guide\\\" target=\\\"_blank\\\" rel=\\\"noopener\\\">Google states that AI Overviews and AI Mode rely on its core Search ranking systems to retrieve pages from its Search index<\/a>, combined with a query fan-out that generates several related searches at once. <a href=\\\"https:\/\/www.perplexity.ai\/hub\/blog\/architecting-and-evaluating-an-ai-first-search-api\\\" target=\\\"_blank\\\" rel=\\\"noopener\\\">Perplexity has published that it runs its own index of more than 200 billion URLs<\/a>, decomposes documents into self-contained spans that are retrieved and ranked individually, and applies cross-encoder rerankers to the final selection. OpenAI says ChatGPT search draws on third-party search providers alongside its own indexed and cached web content.\", \"Three consequences follow for an energy supplier. The engines are reading different indexes, so visibility is not one number. They retrieve passages rather than pages, so a page that is strong overall but vague in its middle sections gets passed over. And the citation you see is a filtered subset of what was consulted: OpenAI's own API documentation notes that the list of sources the model consulted is often larger than the number of citations shown.\", \"There is also a quality problem nobody selling this should hide. The standard academic audit of citation accuracy, <a href=\\\"https:\/\/arxiv.org\/abs\/2304.09848\\\" target=\\\"_blank\\\" rel=\\\"noopener\\\">Liu, Zhang and Liang, published at EMNLP Findings in 2023<\/a>, examined four generative search engines and found only 51.5 percent of generated sentences were fully supported by their cited sources, and 74.5 percent of citations supported the statement they were attached to. Being cited is not the same as being described correctly, which is why we measure both.\"]}, {\"id\": \"the-crawlers\", \"q\": \"Which crawler settings decide whether you appear at all?\", \"h\": \"Training bots and retrieval bots are not the same thing\", \"p\": [\"This is the cheapest and most commonly mishandled part of the whole discipline. Many firms blocked AI crawlers wholesale on legal advice between 2023 and 2025. Some of them removed themselves from AI answers without intending to, and have no idea.\", \"The distinction is between crawlers that gather training data and crawlers that build the retrieval index an assistant reads at answer time. <a href=\\\"https:\/\/developers.openai.com\/api\/docs\/bots\\\" target=\\\"_blank\\\" rel=\\\"noopener\\\">OpenAI documents four separate agents<\/a>: GPTBot for model training, OAI-SearchBot for ChatGPT search eligibility, ChatGPT-User for user-initiated fetches, and OAI-AdsBot. Blocking GPTBot excludes you from future training sets and does nothing to your presence in ChatGPT search. Blocking OAI-SearchBot removes you from ChatGPT search answers.\", \"<a href=\\\"https:\/\/support.claude.com\/en\/articles\/8896518\\\" target=\\\"_blank\\\" rel=\\\"noopener\\\">Anthropic documents three<\/a>: ClaudeBot for training, Claude-User for user-initiated retrieval, and Claude-SearchBot for search indexing. All three obey robots.txt. <a href=\\\"https:\/\/docs.perplexity.ai\/docs\/resources\/perplexity-crawlers\\\" target=\\\"_blank\\\" rel=\\\"noopener\\\">PerplexityBot builds the search index and explicitly does not train models<\/a>, so blocking it removes you from Perplexity results. Google's Google-Extended is a control token rather than a separate crawler, and its documentation now covers grounding in Gemini as well as training, so blocking it can remove you from Gemini answers while leaving Search untouched.\", \"Two further traps. Verify crawlers by the published IP ranges rather than the user-agent string, which is trivially spoofed. And check your content delivery network separately from your robots file: from 15 September 2026 Cloudflare changed its defaults for new customers, new sites of existing customers and existing free-plan customers to block mixed-use AI crawlers on ad-bearing pages. A firm can have a permissive robots.txt and still be invisible because something upstream is returning a 403.\", \"The rendering point closes this section. Log analysis published by Vercel and MERJ found that none of GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot or Meta's agent execute JavaScript. GPTBot fetched JavaScript files in around 11.5 percent of requests and ClaudeBot in around 23.8 percent, but read them as text rather than running them. If your capability statement, specification table or price is built client-side, it does not exist as far as those crawlers are concerned. The test is the page source, not the browser inspector.\"], \"table\": {\"cols\": [\"Agent\", \"What it does\", \"Obeys robots.txt\", \"What blocking it costs you\"], \"rows\": [[\"GPTBot\", \"OpenAI model training\", \"Yes\", \"Exclusion from future training data only. No effect on ChatGPT search\"], [\"OAI-SearchBot\", \"Indexing for ChatGPT search\", \"Yes\", \"Removal from ChatGPT search answers\"], [\"ClaudeBot\", \"Anthropic model training\", \"Yes\", \"Exclusion from future training data only\"], [\"Claude-SearchBot\", \"Anthropic search indexing\", \"Yes\", \"Prevents indexing of your content for Claude search\"], [\"PerplexityBot\", \"Perplexity search index, not training\", \"Yes\", \"Removal from Perplexity results\"], [\"Google-Extended\", \"Control token for Gemini training and grounding\", \"Yes\", \"No Search effect, but can remove you from Gemini answers\"], [\"Googlebot\", \"Google Search index, and the substrate for AI Overviews\", \"Yes\", \"Removal from Google Search and Google's AI surfaces\"]]}}, {\"id\": \"what-works\", \"q\": \"What actually moves AI visibility, on the evidence?\", \"h\": \"Mentions first, structure second, markup last\", \"p\": [\"The strongest measured signal is not on your website. Across 75,000 brands, <a href=\\\"https:\/\/ahrefs.com\/blog\/ai-overview-brand-correlation\/\\\" target=\\\"_blank\\\" rel=\\\"noopener\\\">Ahrefs found branded web mentions correlated with AI Overview presence at 0.664<\/a>, against 0.527 for branded anchors, 0.392 for branded search volume, 0.326 for domain rating and 0.218 for backlinks. Brands in the top quartile by web mentions averaged 169 AI Overview mentions; the next quartile down averaged 14. Ahrefs states plainly that this is correlation rather than proof of causation, and we repeat that caveat rather than burying it.\", \"It is reinforced by a mechanism rather than only a correlation. Research presented at <a href=\\\"https:\/\/aclanthology.org\/2025.emnlp-main.872\/\\\" target=\\\"_blank\\\" rel=\\\"noopener\\\">EMNLP 2025 found that language model citation preference tracks the identity of the outlet rather than the content of the article<\/a>. The models could barely infer an article's stance from its text but recognised the source almost perfectly. If that holds, then who you are is doing more work than what any individual page says, which is an uncomfortable finding for anyone selling page-level optimisation.\", \"On the page itself, two tactics have measured causal support. The benchmark study of generative engine optimisation, <a href=\\\"https:\/\/arxiv.org\/abs\/2311.09735\\\" target=\\\"_blank\\\" rel=\\\"noopener\\\">published at KDD 2024 across 10,000 queries<\/a>, measured adding attributed quotations at plus 43 percent and adding named statistics at plus 33 percent on its visibility metric. Keyword stuffing scored below the baseline. Two caveats that are almost never quoted: the engine was simulated rather than live, and the effect reverses by rank. For a source already ranked first, quotations cost 22.9 percent and statistics cost 20.6 percent. For a source ranked fifth they gain 99.7 percent and 97.9 percent. These are challenger tactics, and most energy suppliers are challengers.\", \"Freshness helps, unevenly. Across 16.975 million cited URLs, AI-cited content was 25.7 percent fresher by publication date, but the effect concentrated in ChatGPT and was essentially absent in Google AI Overviews, which cited content 16 days older than the organic results. Peer-reviewed measurement found the median age of cited articles varies hugely by category, from 62 days for Claude in consumer electronics to 493 days for Google in automotive. Recency is a lever in some categories and noise in others.\"], \"pillars\": [{\"n\": \"01\", \"t\": \"Be fetchable\", \"d\": \"Correct robots rules by crawler, no content delivery network blocking retrieval bots, and substantive content in the server-rendered HTML rather than built by JavaScript after load.\"}, {\"n\": \"02\", \"t\": \"Be mentioned\", \"d\": \"Independent pages that name your firm are the strongest measured correlate. Trade press, association directories, conference listings, industry roundups. This is earned, not marked up.\"}, {\"n\": \"03\", \"t\": \"Be specific\", \"d\": \"You will not out-cite Wikipedia on broad questions. You can win narrow technical questions where no encyclopaedia entry exists, which in energy is most of the questions that matter commercially.\"}, {\"n\": \"04\", \"t\": \"Be checkable\", \"d\": \"Named figures with a source and a date, attributed quotes, explicit methodology. The two tactics with measured causal support are both forms of showing your working.\"}]}, {\"id\": \"what-does-not-work\", \"q\": \"Which popular recommendations do not stand up?\", \"h\": \"The part of this market that is selling folklore\", \"p\": [\"We publish this section because a buyer cannot evaluate a supplier in this field without it, and because most of what is sold under the heading of AI optimisation has no evidence behind it.\", \"Schema markup does not buy citation. A difference-in-differences study tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 against roughly 4,000 matched controls. AI Mode citations moved 2.4 percent and ChatGPT 2.2 percent, both inside noise, and AI Overview citations fell 4.6 percent. <a href=\\\"https:\/\/developers.google.com\/search\/docs\/fundamentals\/ai-optimization-guide\\\" target=\\\"_blank\\\" rel=\\\"noopener\\\">Google's own guidance says structured data is not required and there is no special markup you need to add<\/a>. Schema still earns its place for entity disambiguation and Google rich results. It is not an AI visibility lever and should not be sold as one.\", \"FAQPage markup in particular is now close to worthless. Google removed FAQ rich results from Search entirely on 7 May 2026. The text of a well-written question and answer still helps, because it is visible content that retrieves cleanly. The markup wrapper no longer does anything.\", \"The file called llms.txt is a bet rather than an optimisation. Google's John Mueller has said no AI system currently uses it, and Gary Illyes has said Google does not support it and has no plans to. No major AI company has committed to reading it. A survey of 137,210 domains in May 2026 found 28 percent published one and 97 percent of those received zero requests for it. It costs almost nothing to publish and we do not object to it, but anyone presenting it as a core deliverable is selling you a file nothing reads.\", \"Three more to retire. Rewriting content in an artificially AI-friendly voice: Google states you do not need to write in a specific way for generative search, and the benchmark study's own fluency optimisation sat mid-table and reversed for high-ranked sources. Breaking content into small chunks for AI: Perplexity and Bing chunk internally regardless of what you do, and Google says there is no requirement to do it and no ideal page length. And the widely circulated claim that 87 percent of ChatGPT citations rank in Bing's top 10: the original study covered 100 queries against the top 20 on a superseded product, and larger measurements of 15,000 and 118,931 queries put the figure at 8.1 and 6.82 percent.\"]}, {\"id\": \"energy-specific\", \"q\": \"Why is this different in energy?\", \"h\": \"Narrow questions, gated proof and a procurement layer\", \"p\": [\"Three features of energy change the problem. The first is question shape. General-interest queries are dominated by encyclopaedic and community sources: Wikipedia accounted for 29.7 percent of ChatGPT's top-cited pages in one study and Reddit for 16.8 percent of its citation share. A specialist supplier will not beat those on broad questions. But the questions that decide energy purchases are not broad. What certification does this duty require, which vendors are registered with this operator, how is an in-country value score calculated, what did this award actually include. There is no encyclopaedia entry for those, and whoever answers them properly is the available source.\", \"The second is that the proof lives behind gates. Energy suppliers put their most citable material, the specifications, the certifications, the reference installations, the local content position, into PDFs, portals and sales conversations. Retrieval crawlers read HTML well, PDFs poorly, and gated content not at all. A supplier can be the most qualified firm for a duty and be entirely absent from the machine-readable record of that duty.\", \"The third is the procurement layer. In energy the buyer often is not the payer. The operator writes the specification and maintains the approved vendor list, the contractor places the order, and a local content scheme such as Saudi Aramco's iktva or ADNOC's In-Country Value programme scores the bid. An assistant asked to shortlist suppliers is reasoning across all of that, and the firms that publish their registration status, certification scope and local content position as plain text are the ones it can actually place. We wrote about the underlying structure in <a href=\\\"\/channel-partner-marketing-energy-b2b\/\\\">channel and partner marketing in energy B2B<\/a>.\", \"There is a forward-looking reason to treat this as urgent rather than interesting. Gartner forecasts that by 2028 around 90 percent of B2B buying will be intermediated by AI agents, with more than 15 trillion dollars flowing through agent exchanges, while separately forecasting that by 2030, 75 percent of B2B buyers will prefer sales experiences that prioritise human interaction over AI. Both can be true. The qualification and shortlisting layer becomes machine-mediated while negotiation stays human, which means the machine-readable layer decides who gets into the conversation at all.\"]}, {\"id\": \"measurement\", \"q\": \"How should AI visibility be measured honestly?\", \"h\": \"Grounded and ungrounded are two different numbers\", \"p\": [\"The single most important rule in this discipline, and the one most often broken, is that an API model answering with no web access is not the product your buyer uses. A model can name your firm from training memory while the live consumer product, running retrieval, never surfaces you. Reporting those together produces a flattering number that describes nothing.\", \"Project 54 reports two separate figures. Retrieval visibility is measured with live web access enabled, which is what actually happens when a buyer asks. Model memory is measured without it, which tells you what the model has absorbed about you historically. They move for different reasons and they need different remedies: retrieval visibility responds to being fetchable and being mentioned, model memory responds only slowly and only to sustained presence.\", \"The method is a fixed prompt panel rather than a one-off check, because run-to-run variance is large and any single result is close to meaningless. We fix a set of prompts a real buyer would type, in each language the market actually uses, run them on a fixed schedule from a clean logged-out session with location held constant, and record per run whether the firm was named, whether it was cited with a link, where in the answer it appeared, which competitors were named, and which URLs were cited. Trend is the finding. A single run is an anecdote.\", \"One discipline check that costs nothing and catches a lot: cross-reference the URLs an engine cited against your own server logs. If a page is being cited but was never fetched by that engine's crawler, the citation came from training memory rather than retrieval, and the work required to defend it is completely different from the work required to earn it.\", \"This is what the <a href=\\\"\/findability-audit\/\\\">Energy Findability Audit<\/a> does as a paid engagement, and the measurement discipline above is the product. The reason we publish the method rather than hiding it is that in a field this full of unverifiable claims, the method is the differentiator.\"]}], \"media\": {\"image\": {\"src\": \"\/wp-content\/uploads\/2026\/09\/energy-engineer-inspecting-process-equipment.jpg\", \"label\": \"An engineer inspecting process equipment against a bright yellow structure. The proof that decides an energy purchase, the certification, the duty rating, the registration status, is usually the proof that never makes it into machine-readable text.\", \"credit\": \"Project 54\"}, \"infographicLabel\": \"Measured overlap between an AI assistant's cited domains and Google's top 10: GPT-4o 4.0 percent, Gemini 11.1 percent, Claude 12.6 percent, Perplexity 15.2 percent, from a peer-reviewed study of 1,000 queries across five engines.\", \"pdf\": {\"href\": \"https:\/\/projectfifty4.com\/wp-content\/uploads\/2026\/09\/ai-search-visibility-energy.pdf\", \"title\": \"AI Search Visibility for Energy Companies, briefing deck\", \"meta\": \"PDF\"}}, \"poll\": {\"q\": \"If a buyer asked an assistant to shortlist suppliers for your duty today, what would it find?\", \"options\": [{\"id\": \"nothing\", \"label\": \"Nothing. We have never checked\", \"insight\": \"The most common answer and the cheapest to change. Start with two free checks: whether your robots rules block a retrieval crawler such as OAI-SearchBot or PerplexityBot, and whether your capability content is present in the page source rather than built by JavaScript. Both are answerable in an afternoon and either one can be the whole problem.\"}, {\"id\": \"site\", \"label\": \"Our website, but nothing else about us\", \"insight\": \"This is the position the evidence says is hardest to hold. Branded web mentions correlated with AI presence at 0.664 across 75,000 brands, against 0.218 for backlinks, and research at EMNLP 2025 found citation preference tracks the outlet rather than the article. A site with no independent corroboration is a single point of failure.\"}, {\"id\": \"outdated\", \"label\": \"Something outdated or wrong about us\", \"insight\": \"Worth taking seriously rather than shrugging at. The standard academic audit found only 51.5 percent of generated sentences were fully supported by their cited sources. Misdescription is a distinct failure from invisibility and it needs a different fix, usually correcting the machine-readable record at the source the engine is actually reading.\"}, {\"id\": \"competitors\", \"label\": \"Our competitors, not us\", \"insight\": \"Diagnostic rather than fatal, because it tells you which sources the engine trusts for your category. The useful next step is to look at what it cited, not at what it said. Those pages are the map of where your name needs to appear.\"}], \"note\": \"Answers are anonymous and inform Project 54's ongoing research into AI findability in energy markets.\"}, \"faq\": [{\"q\": \"Is AI search visibility the same as SEO?\", \"a\": \"No, and treating them as one is the most common mistake. They overlap for Google AI Overviews, which run on Google's own index, but barely overlap for the assistants. A peer-reviewed study of 1,000 queries across five engines measured the share of an assistant's cited domains that also appear in Google's top 10 at 4.0 percent for GPT-4o, 11.1 percent for Gemini, 12.6 percent for Claude and 15.2 percent for Perplexity. Good SEO helps you on Google's AI surfaces and does relatively little for the rest. The disciplines share some foundations, mainly crawlability and clear content, and diverge sharply after that.\"}, {\"q\": \"What is the difference between GEO, AEO and AI search visibility?\", \"a\": \"They are largely the same thing under different labels. Generative engine optimisation, or GEO, comes from a 2024 academic paper and refers to increasing a source's prominence in generated answers. Answer engine optimisation, or AEO, is the more common commercial term for the same goal. AI search visibility is the plainest description and the one we prefer, because it names the outcome rather than a technique. Be careful with any supplier who presents these as three separate services with three separate fees.\"}, {\"q\": \"Does blocking AI crawlers protect our content?\", \"a\": \"It depends entirely on which crawler, and the distinction is where most damage is done. Blocking a training crawler such as GPTBot or ClaudeBot keeps your content out of future model training and has no effect on whether you appear in that company's AI search product. Blocking a retrieval crawler such as OAI-SearchBot, Claude-SearchBot or PerplexityBot removes you from those products' answers. Firms that blocked AI bots wholesale on legal advice frequently blocked both without meaning to. It is free to check and takes minutes.\"}, {\"q\": \"Do we need llms.txt?\", \"a\": \"There is currently no credible evidence any production AI system reads it. Google's John Mueller has stated that no AI system uses it, and a survey of 137,210 domains in May 2026 found that 97 percent of published llms.txt files received zero requests. It is cheap and harmless to publish, so we do not argue against it, but it should never displace work on crawler access, server-side rendering or off-site presence. If a supplier's AI visibility proposal leads with llms.txt, ask what evidence they have that anything reads it.\"}, {\"q\": \"How long does it take to improve AI visibility?\", \"a\": \"The technical layer is fast and the entity layer is slow. Crawler access, rendering and content structure are fixable in weeks and their effect appears at the next crawl. Being mentioned across independent sources is a programme rather than a project, and the evidence suggests it is the layer that matters most. Model memory, as distinct from retrieval, lags further still because it depends on training cycles. Anyone promising to move your AI visibility in a fixed short window is either talking only about the technical layer or is not being straight with you.\"}, {\"q\": \"What does Project 54's Energy Findability Audit cover?\", \"a\": \"It measures AI visibility across a fixed prompt panel, in multiple languages, against named competitors, reporting retrieval visibility and model memory as two separate numbers rather than one blended figure. It also covers the technical layer that decides whether you can be retrieved at all: crawler access by agent, rendering, and whether your capability content exists in machine-readable form. It is a fixed-price engagement from 3,500 euros excluding value added tax, with the level set by the number of prompts, engines, competitors and languages covered. Full scope and pricing are on the audit page.\"}], \"related\": [{\"title\": \"The Energy Findability Audit\", \"topic\": \"Services\", \"href\": \"\/findability-audit\/\"}, {\"title\": \"What Is Answer Engine Optimization?\", \"topic\": \"Marketing\", \"href\": \"\/what-is-answer-engine-optimization\/\"}, {\"title\": \"AI Buying Agents and Energy B2B\", \"topic\": \"Marketing\", \"href\": \"\/ai-buying-agents-energy-b2b\/\"}, {\"title\": \"AI Share of Voice for Energy B2B\", \"topic\": \"Marketing\", \"href\": \"\/ai-share-of-voice-energy-b2b\/\"}, {\"title\": \"Channel and Partner Marketing in Energy B2B\", \"topic\": \"Marketing\", \"href\": \"\/channel-partner-marketing-energy-b2b\/\"}, {\"title\": \"Generative Engine Optimization for Energy B2B\", \"topic\": \"Marketing\", \"href\": \"\/generative-engine-optimization-energy-b2b\/\"}], \"newsletter\": {\"kicker\": \"The Energy Growth Brief\", \"title\": [\"Intelligence,\", \"to your inbox\"], \"body\": \"Join energy and industrial leaders getting our marketing, AI-growth and revenue-architecture intelligence, direct, no filler.\", \"placeholder\": \"you@company.com\", \"cta\": \"Subscribe\", \"note\": \"No spam. Unsubscribe anytime. We read every reply.\"}}","p54_faq":"","p54_media":"","p54_comments_enabled":"","footnotes":""},"categories":[92,125],"tags":[],"class_list":["post-4209","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-analysis","category-strategy"],"acf":[],"_links":{"self":[{"href":"https:\/\/projectfifty4.com\/fr\/wp-json\/wp\/v2\/posts\/4209","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/projectfifty4.com\/fr\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/projectfifty4.com\/fr\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/projectfifty4.com\/fr\/wp-json\/wp\/v2\/users\/12"}],"replies":[{"embeddable":true,"href":"https:\/\/projectfifty4.com\/fr\/wp-json\/wp\/v2\/comments?post=4209"}],"version-history":[{"count":1,"href":"https:\/\/projectfifty4.com\/fr\/wp-json\/wp\/v2\/posts\/4209\/revisions"}],"predecessor-version":[{"id":4216,"href":"https:\/\/projectfifty4.com\/fr\/wp-json\/wp\/v2\/posts\/4209\/revisions\/4216"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/projectfifty4.com\/fr\/wp-json\/wp\/v2\/media\/4205"}],"wp:attachment":[{"href":"https:\/\/projectfifty4.com\/fr\/wp-json\/wp\/v2\/media?parent=4209"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/projectfifty4.com\/fr\/wp-json\/wp\/v2\/categories?post=4209"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/projectfifty4.com\/fr\/wp-json\/wp\/v2\/tags?post=4209"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}