Your Search Impressions Are Not Buyer Demand

The definition is the whole problem

Google's documentation is unambiguous and it is not what most marketing reports assume. An impression means that a user has seen (or potentially seen) a link to your site in Search, Discover, or News. In general, an impression is counted whenever an item appears in the current page of results, whether or not the item is scrolled into view, as long as the user need not click to see more results.

Read that carefully. The unit of measurement is appearance on a results page, not perception by a person. The parenthetical, potentially seen, is doing an enormous amount of work. A request issued by a rank tracker produces a results page. A request issued by an AI assistant produces a results page. Neither involves a human who could become a customer, and neither is distinguishable from one in the data Google returns to you.

Google has never claimed otherwise. The same help page states that the heuristics are subject to change and contains no reference anywhere to filtering bot, automated or non-human traffic. The only on-record position from Google comes from John Mueller in August 2021, responding to a site owner who found query strings exactly matching their page titles accounting for a fifth of monthly clicks. His two replies were: sometimes it can be from bots, we don't necessarily filter all of that out in Search Console, and, we do try to filter these out in Search Console, but it is sometimes visible. Those two statements sit in mild tension and Google has not reconciled them. They remain the clearest available answer, and the answer is that filtering is attempted and incomplete.

This matters more in energy B2B than in most sectors, because the absolute volumes are small. A consumer brand with ten million monthly impressions can absorb a few thousand machine-generated ones inside the noise floor. An industrial supplier with ninety thousand a quarter cannot. At our scale a single scheduled query is visible in the quarterly total, and we can prove it.

المشروع 54A dense instrument manifold reads pressure at many points and tells you nothing unless you know which gauge is connected to what. Search Console is the same instrument.

September 2025 ran the experiment for everyone

On 12 September 2025 Google disabled support for the num=100 URL parameter, which returned 100 results in a single request and was the standard mechanism rank trackers used to read deep positions cheaply. A Google spokesperson told Search Engine Land only that the use of this URL parameter is not something that we formally support, declining to say whether the change was a bug or intentional, why it was made, or what the Search Console impact would be.

The impact was measured independently within days, and collated by Brodie Clark. The figures below are the closest thing the industry has to a controlled experiment on how much of an impression count is machinery.

Clark's mechanism explanation is the part worth keeping: for a normal search result for a typical Google user, the default current page would be 10 organic listings. Because rank trackers tend to use 100 organic listings as the default view, a result that ranks in position 99 would then incur an impression within GSC, even though it is a bot impression. On one property he observed a decline of more than 200,000 daily desktop impressions, concentrated on desktop, which is what rank trackers default to. His own summary was blunt: our GSC data has unfortunately been polluted with a lot of bot impressions over the past year, which has reduced the reliability of the dataset.

Two cautions before anyone treats this as settled. First, these are practitioner datasets of varying size and method, not peer-reviewed work, and the range between them, from a 15 per cent median to a 25 per cent aggregate, is wide. Second, num=100 was one mechanism among several. Closing it removed deep-position scraping. It did not remove scheduled queries that read only the first page, which is what most brand monitoring platforms do, and it did nothing about AI assistants.

Signatures you can actually find

Every example in this section is from Project 54's own Search Console, URL prefix property, for queries between 29 June and 25 September 2026. We are publishing our own numbers rather than a client's because the point is only credible if the person making it is willing to show the inside of their own reporting.

Class one is monitoring and rank tracking software. Its signature is search operator syntax, which no buyer types. Our largest example is a quoted phrase, shell global, paired with a strategy term: 520 impressions, zero clicks, average position 9.2. A second variant of the same quoted phrase took 218 impressions, also zero clicks. A third took 43. Six operator variants of one document title, growing shareholder value: a reset bp, collectively took 365 impressions and returned nothing. The most conclusive single example is a query consisting of roughly twenty negative site exclusions wrapped around the phrase measure share of voice. That is a media monitoring platform's default query template, logged verbatim as a search term.

Class two is AI assistants. Google's own documentation confirms the mechanism: for AI Mode, if a user asks a follow up question within AI Mode, they are essentially performing a new query, and all impression, position and click data in the new response are counted as coming from this new user query. AI Overview links are also counted, with all links in an overview assigned the same single position. The result is that conversational fragments appear as search queries. Ours include yes at 79 impressions and average position 6.1, both at 9 impressions, the German beides at 3 impressions and average position 1.0, and the Romanian a doua at position 1.0. No human types beides into Google and lands on an energy marketing site at position one. These are follow up turns inside an answer.

The same class produces the opposite extreme: whole briefs. One of ours reads, are the major us wellhead/christmas tree specification and the corresponding iso 10423 document technically equivalent, or are there real clause-level differences that matter for procurement language. Another asks which platform combines lead enrichment with workflow orchestration. A third asks which of the American Petroleum Institute, the American Gas Association and AFPM produces the highest quality statistical output, explaining that the asker is setting sourcing guidance for a program office. These are not searches. They are questions put to an assistant, which then queried Google. Suganthan Mohanadasan's four signal test catches them precisely: an instruction to an assistant, first person context, a dangling pronoun, or politeness, at ten words or more.

Class three is the single saved query, and it is the one that does the most damage to a quarterly report. Our largest query by impressions in the period is china fuel oil inventory february 2026: 2,671 impressions, zero clicks, average position 8.1. A six word string that specific does not have two and a half thousand distinct searchers in a quarter. Mohanadasan's discriminator is repetition: conversational and hyper specific strings almost never repeat, because no two people phrase the same need identically. A string that repeats at volume is one scheduled process. The tell is not the position and not the click rate, it is the concentration.

Our worst case is at page level rather than query level. The page energy asset acquisition risk assessment framework for buyers took 8,448 impressions at average position 2.0 and returned 2 clicks, a rate of 0.02 per cent, from essentially one query which is that page's former title, word for word, more than a month after the page was retitled. Position 2.0 on a title nobody uses any more, with no clicks, is not demand. It is a saved alert in somebody's monitoring tool. We have now confirmed this four times across four consecutive audits and we are retiring it from our watch list.

The audit, with its limits stated

We classified every query Google was willing to show us. The rule set was deliberately conservative: operator or quoted syntax, or a bare conversational fragment, or ten words or more with an instruction or question structure. Anything else was left in the human column, including the 2,671 impression saved query, because the classifier cannot prove concentration from a single row.

The result: 218 distinct queries, 2,432 impressions, 2 clicks. Adding the two cases the rule set deliberately misses, the single saved query at 2,671 and the retitled page at 8,448, gives 13,551 impressions and 4 clicks.

Against site totals for the 89 days to 27 September 2026 of 93,496 impressions and 607 clicks, that is 14.5 per cent of impressions and 0.7 per cent of clicks. The query level window runs to 25 September rather than 27, a two day mismatch that does not move the conclusion.

Now the honest limits, because a number like that is only useful if you say what it cannot do. It is a floor, not an estimate. It counts only patterns we could positively identify, and the tail is full of two and three word strings that could be either. It cannot see most of the data at all: Google anonymises a large share of queries for privacy, and the rows it returned to us account for well under half of total impressions. Kevin Indig, analysing 450 million impressions, headlines the problem as Google filtering three quarters of search impressions. On one practitioner's own BigQuery export over 59 days, 57.7 per cent of web impressions carried no query string at all. So the correct statement is not that 14.5 per cent of our impressions are machinery. It is that at least 14.5 per cent are, the true figure is higher, and a material share is permanently unauditable.

Benchmark the SERP, not the position

A 0.65 per cent site CTR at average position 6.9 looks like a crisis. It is not, and the correction matters because the wrong diagnosis leads to the wrong work.

Advanced Web Ranking's July 2026 snapshot, built on more than 1.75 million US SERPs a month, puts position 7 at 0.46 per cent across all SERPs and 0.35 per cent where an AI Overview is present. At position 6 the figures are 0.73 and 0.55 per cent. Our 0.65 per cent at position 6.9 is at or above the benchmark for where we rank. There is nothing to fix in the click rate.

What has genuinely changed is the shape of the results page. Advanced Web Ranking found AI Overviews on 43 per cent of sampled desktop SERPs, 74.3 per cent zero click behaviour on those pages against 53.6 per cent without, and only 26 per cent of desktop SERPs now purely organic. Position 1 averaged 20.9 per cent over the last twelve months against 28.0 per cent the prior twelve, a 25 per cent fall year on year. Their conclusion is the sentence every energy marketer should have pinned above the dashboard: your CTR benchmark is a property of the SERP, not of your ranking. A 12 per cent CTR at position 1 is excellent on an AI Overview SERP and underwhelming on a clean organic one.

Treat those figures with the caveats their authors give. Advanced Web Ranking's numbers are modelled click and impression estimates per position, not measured site data, they are impression weighted so high volume queries dominate, and their own month to month series swings violently, from 25.58 per cent in January 2026 to 1.96 per cent in April to 9.69 per cent in July for position 1 with an AI Overview. Any single month is a weak benchmark. Ahrefs, using 300,000 keywords and aggregated Search Console data, put the AI Overview effect at position 1 at minus 58 per cent in December 2025, up from minus 34.5 per cent in their April 2025 study. Advanced Web Ranking's equivalent is minus 52 per cent. Two credible studies, two different methods, a six point gap. Cite whichever you like, but cite one, and never average them.

The wider backdrop, from Rand Fishkin at SparkToro: 68.01 per cent of US Google searches ended without a click between January and April 2026, up from 60.45 per cent in 2024, which he calls the fastest acceleration of this phenomenon in the last decade. Notably, only 0.34 per cent of searches reached AI Mode in that window, so AI Mode is not yet the cause. Fishkin is explicit that his panels changed between periods and are not perfectly comparable.

A reporting discipline, not a tool

None of this argues for ignoring Search Console. It argues for reporting it honestly, which in a small volume industrial market means doing four things.

First, quantify the deduction rather than debating it. Run the classification once a quarter, publish the figure, and show the method. A board that has been shown the deduction will trust the remainder. A board that later discovers the deduction will not trust anything.

Second, do not report impressions as a headline at all. Report qualified sessions and pipeline, and keep impressions as a diagnostic. At energy B2B volumes, percentage metrics on small denominators are unstable by construction, and a single scheduled query can move a quarterly percentage by a point.

Third, rule out the benign explanations before blaming machines. Brodie Clark's seven causes of a low click impression spike, internal sitelinks, image packs, knowledge panel images, tweet carousels, Web Stories, ranking tests and AI Overviews, cover most cases, and six of the seven cannot be isolated in the Search Appearance filter, which is why they confuse people. His own caution is worth repeating: people do like to blame the spike in impressions without clicks on AI Overviews, but there is just as much likelihood that features as simple as images in knowledge panels or image packs are the greater contributor.

Fourth, treat the conversational queries as intelligence rather than noise, even though you exclude them from demand. The wellhead specification question in our data tells us a procurement team is comparing API and ISO clause language and cannot get a clean answer. That is a content brief, delivered free, from a buyer we will never otherwise meet. It belongs in the editorial plan and out of the demand number. Mohanadasan's rule is the right one: exclude the machine buckets, the probes and harnesses, from any opportunity analysis, and treat artefacts as a signal rather than demand.

The forward view is that this gets worse before it gets better. Google rolled out a generative AI performance report to all websites worldwide on 31 August 2026, but it reports impressions, pages, countries, devices and dates, with no queries and no clicks, and inherits the same 1,000 row limit. More of your visibility is moving into surfaces that report less about themselves. The teams that will be trusted in 2027 are the ones that started publishing the deduction in 2026.