The Prompt-Shaped Queries in Your Search Console Are Mostly Other People's Monitoring Tools

The Prompt-Shaped Queries in Your Search Console Are Mostly Other People’s Monitoring Tools

By Stephen Honight, Founder of Lmo7

There is a query in Lmo7’s own Search Console that reads “compare lmo7 vs other ai commerce optimization agencies”. It has an average position of 1.0. It has zero clicks. Nobody typed it into a search box. Somebody’s tool did.

That one line is the whole story of this piece. A growing share of the long, prompt-shaped queries showing up in Search Console are not buyers phrasing questions. They are AI visibility platforms firing monitoring prompts at Google surfaces, and Search Console is recording the traffic. If you are building content plans or client reports off those queries without filtering them, you are measuring machines and calling it demand.

The short version, before the detail. The popular technique of mining long Search Console queries for buyer language is sound, but the input is now contaminated by competitor monitoring tools. The commercial cost of missing this is real: query themes that are actually a vendor’s monitoring schedule end up shaping content budgets and getting reported to boards as demand. The fix takes about an hour. Separate the two piles with a regex and a handful of tells. The human pile feeds your prompt tracking and content plan. The synthetic pile, read properly, tells you which competitor set you have been placed in, what criteria you are being judged on and which third-party sources actually settle the comparison. Both piles are useful. Confusing them is the expensive mistake.

The accepted reading, fairly put

Chris Long set out the original technique in Search Engine Land on 27 February 2026. Filter your Search Console queries to ten words or more using the regex ^(?:\S+\s+){9,}\S+$ and what comes back reads like prompts rather than keywords. His argument is that these long queries are a window into how real people phrase questions, and the practical use is to build your tracked prompt set from them rather than guessing what buyers might ask.

That guessing problem is real. Rand Fishkin, cited in the same article, found that when 142 respondents were asked to write a prompt for the same task, the similarity between their prompts was 0.081. In plain terms, you will never guess the exact prompt a buyer uses, so observed query data is genuinely valuable. Long is also careful about what he does not claim: there is no direct proof the queries come from ChatGPT or AI Mode, and some may simply be people using Google like a chatbot.

The technique holds up. What has changed is the input.

What the Trakkr data does to that reading

Trakkr, an AI visibility tool vendor, published a read of their own Search Console property in early September. Their figures, which are theirs alone, self-reported from a single property and not independently verified by Lmo7: over 16 months, 3,476 structured queries produced 61,966 impressions and zero clicks. Forced-ranking prompts alone accounted for 44,629 of those impressions. One query family, variations of “what do users say about HubSpot AEO vs Profound vs Scrunch for AI brand visibility tracking on G2?”, runs at over a thousand impressions a month. Google associated 94.8% of the comparison impressions with review pages.

The queries fall into recognisable formats. Forced best-to-worst rankings. “Evaluate this brand for this use case.” “Give a definitive answer with pros and cons.” Persona-conditioned questions where the asker claims a job title and a company size. The same prompt re-run across countries and languages. Some even carry raw configuration values that were never cleaned up, strings like enterprise_1000_5000 dropped into otherwise readable sentences. A human does not write snake_case into a question. A template does.

Trakkr go further and attribute fingerprints to specific vendors: persona cycling that resembles Profound’s public persona setup, a location wrapper they attribute to Bright Data containing the string “not for language”, and daily comparison plus review-source prompts resembling HubSpot and XFunnel’s published AEO programme. I could not verify the Bright Data detail against public documentation, so treat that as Trakkr’s observation rather than established fact.

The surrounding facts check out independently, though. HubSpot acquired XFunnel on 31 October 2025 and launched HubSpot AEO in beta on 14 April 2026, running tracked prompts daily across ChatGPT, Gemini and Perplexity. Daily automated comparison prompts are a documented product behaviour, not a theory. Profound markets persona-segmented prompt tracking as a feature, including region and audience dimensions. And AI queries leaking into Search Console has precedent: Jason Packer reported ChatGPT queries appearing in Search Console in November 2025 and OpenAI confirmed the leak.

Why this lands in Search Console at all

The mechanism is simple once you see it. AI visibility platforms run their tracked prompts through the surfaces they monitor, and at least some of those surfaces are Google’s, including AI Mode. Google announced generative AI performance reporting for Search Console on 3 June 2026, first with a UK test group and worldwide by the end of August, so impressions from AI surfaces now flow into the same property data marketers have always mined.

Monitoring prompts have a distinctive statistical signature in that data. They are zero-click by construction, because a measurement tool never clicks through. They tend to sit at flattering average positions, because a prompt that names your brand returns your brand near the top. And they repeat on a schedule, because the vendor’s product refreshes daily. Stack enough of them into a property and any impressions-based visibility number inflates, any average-position number flatters and any query-theme analysis starts describing a machine’s model of your market rather than your market.

This is the same argument we made in last week’s piece on measuring Rufus honestly. The number is only as good as what went into it. Here is a live case of a measurement input being wrong for a reason almost nobody was checking for.

How to tell the two piles apart

You can run this on your own property in under an hour. It needs Search Console access and a spreadsheet, nothing else.

1. Open Search Console, Performance, Queries. Add a filter: Query, Custom regex, ^(?:\S+\s+){9,}\S+$. That isolates queries of ten words or more.
2. Export the result and sort by impressions.
3. Flag a query as probably synthetic where two or more of these hold: zero clicks at an average position under 5; it names two or more competitor brands; it names a review or comparison source such as G2, Trustpilot or Reddit; it contains an underscore, a snake_case token or a raw country or company-size code; it is phrased as an instruction (“rank these”, “give a definitive answer”) rather than a question; it repeats near-identically across countries or languages.
4. Flag a query as probably human where it is messy, misspelled or incomplete, or where it sits at a deep average position with zero clicks. Real people write badly and rank nowhere in particular.
5. Split the export in two. The human pile feeds your tracked prompt set and your content plan. The synthetic pile feeds a competitor and criteria read, which we will come to.
6. Re-run it monthly. A new entrant in the synthetic pile means a new vendor has added you to somebody’s comparison set.

One distinction worth being precise about. Zero clicks on its own proves nothing. In the same Lmo7 pull, “AI search audit” returned 411 impressions with zero clicks at average position 60.9. That is an ordinary long-tail keyword sitting too deep to earn a click, and the fix is ordinary content work. Zero clicks at position 1.0 on a sentence no human writes is a different animal entirely.

The discard pile is competitive intelligence somebody paid for

Hygiene is the first job, and the obvious place to stop. My sense is the more interesting value sits in the pile you just filtered out.

Those synthetic prompts are a vendor’s working model of your market, and it has leaked into your reporting for free. Read them properly and you can see the competitor set someone else has decided you belong in, the evaluation criteria being forced into every comparison, the personas being simulated against you down to seniority, company size and country, and which markets and languages are being run.

The commercially useful line in Trakkr’s data is the last one: Google associated 94.8% of the comparison impressions with review pages. If comparison questions get answered from review surfaces, then review and third-party comparison content is what settles the answer. That converts a vague instruction like “build authority” into a named target list. You know which sources the engines pull for your category’s comparisons, because the monitoring prompts just showed you.

The consumer-brand version

Lmo7’s readers are mostly consumer brands, not SaaS vendors watching G2. The surfaces change but the mechanism carries straight across. For a consumer brand the equivalents are retailer pages, review sites, category listicles and Amazon’s own review and Q&A content. When a shopper, or an AI assistant acting for one, asks a comparison question about protein bars or pain relief gel, the answer gets assembled from third-party evidence. Brand-owned copy alone does not decide the outcome.

This is the two-track model we keep coming back to, arriving from a new direction. The quick track is your own content and PDP structure, the things you control and can fix this quarter. The long track is review density, retailer presence and third-party comparison content, the things that decide whether the engines trust what you say. The synthetic pile in your Search Console just handed you the target list for the long track. Run both tracks in sequence, because either one alone leaves the job half done.

Our own property has the same pattern

To check this was not exotic, we pulled Lmo7’s own Search Console data via Ahrefs on 1 September 2026, covering May 2025 through August 2026. The property returned 67 query rows in total. Five of them are prompt-shaped rather than search-shaped:

That is 31 impressions, zero clicks, with three of the five sitting at average position 1.0. Look at the shapes: a review-source request, a comparison-set request, a call to action phrased as a query. To be honest, I find the middle one quite funny. Somebody’s tool is out there evaluating whether to recommend us, which means somebody configured a comparison set with Lmo7 in it.

The honest framing: this is five queries on a small, low-authority property, against Trakkr’s 3,476 on theirs. It is a demonstration, not a study. The point is that a reader can find their own version of this in an afternoon, at whatever scale their property runs at.

What I would do next

The takeaway is to clean the input and then go win the surfaces the comparisons are settled on. Nothing here says stop tracking AI visibility. Tracking is how you catch this class of problem in the first place.

1. This week, in-house. Run the ten-word regex on your own property, split the export into human and synthetic piles using the tells above, and check which pile your current reporting themes came from. Owner: whoever holds Search Console. The commercial rationale: it tells you whether the query themes shaping your content budget are demand or somebody else’s monitoring schedule, before the next planning cycle locks in.

2. Next 30 days. Rebuild your tracked prompt set from the human pile only, and stand up the synthetic pile as a monthly competitor and criteria read. If your team wants the data layer and will run the work themselves, that is DaaS at £250 + VAT per month per marketplace. If you want it run for you, it sits in Challenger’s Agentic Tracking and Content Optimisation module at £1,000 per month.

3. Next quarter. Go after the sources the comparison prompts attach to: review density, retailer and marketplace content, third-party comparison pages, Amazon review and Q&A depth. That is Challenger territory, Amazon Content Optimisation at £750 alongside Agentic Tracking and Content Optimisation at £1,000, or the Full Agentic Stack at £2,500 where both tracks run together. For multi-brand portfolios that need the same read across several brands and markets, the Enterprise shape starts from £5,000.

My recommendation would be to start with step one this week even if nothing else moves. It costs an hour and it tells you whether the numbers you are already reporting can be trusted.

Stephen Honight is the founder of Lmo7, a UK agency helping consumer brands win visibility in AI-powered search and agentic commerce, across Amazon and D2C. Lmo7 works with brands including Trip Drinks, Veloforte, Brown-Forman, Haleon, Pelotan and Symprove.