A 100-Word Buyer Brief Ranked at Position 4.1: What Search Console Shows About Prompt-Shaped Demand

A 100-word Tinuiti vs Dentsu evaluation brief ranked at position 4.1 in our Search Console. It read less like a keyword than a buyer handing an assistant a job: compare these agencies against a set of requirements. Google had already logged the question for us.

That is the useful part of the prompt-discovery argument. You do not need to buy a feed to learn every question buyers ask in AI tools before you can improve a page. Our own Search Console shows paragraph-length, persona-stuffed queries, complete with budgets and comparison criteria. They may be typed by people, rewritten by assistants, or both. Search Console does not tell us which. It does tell us where our pages appear for those questions, and where an answer may be missing.

I read 60 queries from our account to see what that log could show. The result is a way to find content gaps, not a count of ChatGPT citations. Here is the data, the boundary around it, and the filter I would run before paying for prompt discovery.

The sample: 60 queries close enough to inspect

I started with a three-month, non-brand striking-distance cut: average positions 4 through 15, then worked down the queries with the most impressions. That is a common way to find terms where a better answer may matter (the striking-distance approach). I stopped at 60 so I could read each row. This was a hand-classified sample from our property, not a measurement of every query in Search Console or every question asked in an AI product.

The categories were deliberately plain:

These are reading labels, not Google classifications. A long query is not automatically an AI prompt, and a short one is not automatically human. The labels help me decide what to do with a row once I can see it.

One boundary matters before looking at the examples: the Tinuiti vs Wpromote comparison at position 3.1 and 111 impressions is an adjacent observation from our account, not a member of the positions-4-to-15 sample. I include it because it shows the same query shape just above the cutoff. The Dentsu brief at position 4.1 and the $4k-a-month agency-shopping query at position 5.3 sit inside the striking-distance frame. All three figures come from our Search Console rows; they are not public benchmarks.

What stood out in the rows

The short keywords were still the majority of the rows I reviewed. I am not claiming that prompts took over search. The striking part was the amount of decision-making context packed into individual long queries: named agencies, a budget, a business type, a fee model or a request for a head-to-head judgment. Those details make a row useful even when it is not the most frequent query on the list.

The agency comparisons were especially instructive. Someone searching ppc agency could be at the beginning of a search. A paragraph comparing named agencies against a budget and fee model asks for a much narrower answer. Our comparison pages appeared for some of those requests because they named the agencies and discussed how the fee structures differed. I used to tell clients to avoid naming competitors on their own site. I was wrong about the blanket rule. If a page will not address the comparison, it cannot answer that comparison directly.

That does not mean every long row deserves a new page. Position is an average, impressions are exposure rather than sales, and Search Console does not reveal the person or process behind a query. But a row at position 4.1 tells me something actionable: Google has already considered one of our pages relevant to a detailed buyer question. I can inspect that page and see whether it answers the question it is appearing for. The query gives me the brief; it does not give me the verdict.

What the log can and cannot say about AI

Long, multi-clause queries resemble the requests people write in chat interfaces. Others have documented that shape in Search Console (examples of long queries). It is tempting to call each one a prompt leaked from ChatGPT. I would not put that claim in a client report.

A Search Console row is the query Google reports for an impression on your property. It is not a private-chat transcript or proof that an assistant cited your page. A person could have pasted a long brief into Google. An assistant could have rewritten a request before using a search provider; OpenAI describes that possibility, but the row alone cannot identify its origin (why origin is hard to establish). A regex can find a shape, not authenticate its author (what long-query filters can and cannot show).

Near-zero clicks would not settle it either. A searcher might read a result without clicking; a query might surface in an AI-mediated search. Search Console does not label those possibilities on the row. The useful inference is smaller and sturdier: our site earns impressions for detailed questions that our content may not yet answer well. Whether a human or a tool submitted a particular one does not change the editorial work.

A 20-minute check for questions your pages nearly answer

Here is the test I would run on your property. Question: which detailed buyer requests already earn impressions, but land on a page that fails to address their constraints? My expectation is that a non-brand export will contain at least some rows that read like tasks rather than traditional keywords. I would not assume a particular share or click-through rate for your site before seeing it.

Keep the controls simple: use one three-month window, exclude your brand, sort by impressions, and keep the same window and filters when you check again. Search Console’s query filters can narrow the text. Filter average position in the exported sheet, rather than treating a position range as a Search Console query-filter setting. That gives you the positions-4-to-15 frame used here without hiding how the cut was made.

1. Open Performance → Search results in Search Console. Set the three-month window, exclude brand terms and export the queries with impressions, clicks, CTR and average position. Sort by impressions; in the sheet, keep rows with average position greater than 3.9 and smaller than 15.1.
2. To inspect long questions in the interface, try a Query → Matches regex filter: (?i)^(what|which|who|how|why|should|can|is|are|does|do)\b.{30,}. The {30,} means at least 30 characters after the opening word, not 30 words, so read the matches before calling them prompts (the long-query recipe). Keep your unfiltered export too; otherwise you will miss prompts that start with a business description rather than a question word.
3. If the question-starter filter catches little, inspect other shapes: a word-count regex such as ^(?:\S+\s+){9,}\S+$, or queries containing comparison language such as best|vs|alternatives. Task verbs such as write|draft|compare|estimate|give me|can you are another way in. These are alternative screens, not proof of AI origin (the filter options and RE2 notes).
4. Read the top 60 eligible rows by hand. Tag each as keyword, question or persona-prompt. Beside each, record its impressions, average position, CTR and the page that appears for it. Then open that page and check whether its first useful answer addresses the query’s actual requirements.

This setup has a limit worth keeping visible. Sorting by impressions helps you find questions with measurable exposure; it does not give you a representative sample of everything buyers ask. A low-impression row may still describe a valuable sale. The first pass is for choosing work, not estimating the size of an entire market.

Sort the result into gaps, near-misses and noise

Once the rows are tagged, I use three working labels. They describe what to inspect next, not what caused the impression.

The $4k-a-month agency-shopping query at position 5.3 illustrated a gap before a pricing comparison section addressed its budget constraint. The Tinuiti vs Wpromote row, at 111 impressions and position 3.1, illustrated a near-miss: the page appeared, but its old opening hedged instead of getting to the fee comparison. Those diagnoses come from reading our rows alongside our pages. The numbers alone cannot tell you whether the copy is good.

To order the work, I use a rough click-opportunity estimate: impressions × (estimated CTR at a target position − current CTR). The striking-distance method uses illustrative CTRs of about 10 percent at position 3, 6 percent at position 5, 3 percent at position 8 and 1.5 percent at position 12 (the method and assumptions). This is a prioritization aid, not a forecast. If two rows both have 400 impressions and 1 percent CTR, they receive the same estimate when I give them the same target CTR, regardless of whether one currently sits at position 5 and the other at 8. Their starting positions still matter when I judge how plausible the climb is.

I would fix the highest-value clear gap first, then the near-miss whose page already has the right subject but buries its answer. I would leave the one-off noise alone. That is a more useful sequence than writing a page for every query with an unusually long string.

Where a tracker earns its fee

Search Console answers the discovery question: which queries earned an impression for our property? It does not answer the monitoring question: how often does an AI product mention or cite us for a fixed set of buyer prompts? Those are different jobs. Reports about generative-AI visibility do not supply the query-level link you would need to turn one Search Console row into proof of a citation (the reporting limit). If the missing citation is the problem you need to diagnose, start with why ChatGPT doesn’t mention your brand, not with an invented interpretation of a Google impression.

Repeated measurement matters because answers move. In one large test, 2,961 recommendation prompts were run 60 to 100 times each; the reported chance of receiving the same brand list twice was under 1 in 100 for ChatGPT or Google AI, with the same list in the same order rarer still (the variability test). That does not tell you what will happen for your prompts. It does tell you why a screenshot of one answer is a poor baseline.

A tracker has a defined use: submit the same buyer-prompt set to the products you care about on a schedule, then watch mentions, cited URLs and competitor share across runs. It has a cost. The draft price board lists Semrush from $99 a month per domain, Ahrefs Brand Radar from $199 a month per index plus a base plan, Profound’s ChatGPT-only Starter at $99 a month on yearly billing, Peec at $95 a month for 50 prompts across three of six models, and Otterly at $29 a month for 15 prompts (the price board). Check the plans before buying; these figures are a comparison from that board, not part of our Search Console analysis.

My order of operations is uncomplicated: use the free log to find candidate questions. Pay for repeated prompt monitoring when you need a weekly mention-share number across models. Do not buy the second job because a sales page tells you the first is impossible without it.

The edit the numbers justify

I would take the top three persona-prompts from the review and give their constraints a place on the relevant pages. A gap needs a direct section that addresses the named competitors, fee model, budget and verdict when those are what the query asks for. A near-miss needs its opening rewritten so the answer arrives before the throat-clearing. On our Tinuiti vs Wpromote page, that means leading with the fee difference rather than making the reader hunt for it. I would not claim that change wins an AI citation; Search Console cannot verify one.

After publishing, I would pull the same filter again after 14 days and compare the rows with the original window in mind. If position holds and CTR rises, the page may be answering the Google query more effectively. If impressions rise and CTR stays flat, I would inspect the snippet and page again before deciding the question needs its own home. Neither outcome proves what ChatGPT did. Both can guide the next edit.

The decision is narrower than the prompt-discovery pitch, and more useful. Our log already contained a 100-word buyer brief at position 4.1, a budget-specific agency question at 5.3 and a related comparison at 3.1. I still see a reason to pay for a tracker when someone needs repeated mention-share reporting across models. For finding the questions to answer next, I start with Search Console. The buyer’s requirements are sitting in the rows. Read them, answer them directly, and make the page earn the impression it already gets.