Ground Your AI Prompt Set in Search Console

01The short answer

To choose the prompts worth tracking for AI visibility, export your Google Search Console queries for the last month, sort them by impressions rather than by clicks or average position, take the top 25, and rewrite each one as the question a buyer would actually type into an assistant. Then prune by consequence rather than by volume.

The sorting step is the whole argument. Sorting by impressions keeps the queries where buyers are looking for something you sell and not finding you, which is precisely the population an answer engine optimization programme exists to change. Sorting by clicks or by position, which is what every published method does either explicitly or as a side effect, keeps the queries where you are already winning and quietly deletes the rest.

The rest of this piece shows what the published methods actually say, why the filter they share is the wrong one, and a worked example using our own Search Console numbers, including the query that every one of those methods would have thrown away.

02What every guide actually tells you

There is no shortage of guidance on choosing prompts. There is a remarkable amount of agreement in it, and the agreement is the problem. Here is what the main published methods say, read from the source rather than from each other.

The last row is the interesting one, because that method starts almost exactly where this article does. It tells you to export everything and warns you off pre-filtering in as many words: do not pre-filter to the queries you already like, because the volume is what makes the pattern reliable. It sets sensible thresholds, at least 1,000 impressions across the export and at least 100 distinct queries, below which it says the patterns are too thin to trust. And then its final step throws the benefit away by keeping only the prompts where your domain already shows up in the answer.

That diagnosis is correct and worth repeating. Most prompt sets in this category are invented, and an invented prompt set produces a dashboard that measures an imaginary market. The disagreement is only about what you do once you have the grounded list.

03Why that is backwards

Filtering a prompt set toward queries you already rank for, or already appear in, guarantees a measurement that cannot find what you are looking for. It is worth being precise about why, because the mistake is intuitive rather than careless.

The intuition is reasonable: a query you rank for is a query where you are relevant, so it is a query where an answer engine might plausibly name you, so it is a good candidate for tracking. Every step of that is true. The problem is what it implies about the queries it excludes. A query with heavy impressions and no clicks is not a query where you are irrelevant. It is a query where Google already believes you are relevant enough to show, and buyers already care enough to search, and you are nonetheless not the answer. That is not noise. That is the exact population an AEO programme is built to move.

There is a second reason, specific to AI answers rather than to search. An answer engine names roughly three brands. A search results page has around ten organic slots plus a long tail of onward navigation. Being eleventh in search is a bad day; being fourth in an answer is being absent. That makes the population of queries where you are visible but not chosen far larger in AI answers than in search, and it is a population that a click-ranked prompt set is structurally blind to.

A third reason is commercial. The stated goal of AEO for a B2B SaaS company is to be present when a buyer is choosing. The queries where you already win are, by definition, the ones where that is already happening. Spending your measurement budget there tells you how well you are defending, and nothing at all about where you are losing.

04A worked example, with real numbers

This is easier to see in first-party data than in the abstract, so here is ours. These are real figures from our own Google Search Console property for linkeddit.com, covering 3 June to 29 August 2026, with bot-shaped queries filtered out.

The query is ai citation tracking. It carries 892 impressions at an average position of 27.6, and it has produced no clicks at all. Its immediate family carries another 490: ai citation finder at 162 impressions and position 35.5, ai citation analysis at 160 and 35.5, ai citation tracker at 104 and 25.1, and ai citation monitoring tool at 64 and 28.6. None of them produced a click either.

Now apply each published method to that data. Ahrefs method one keeps questions the site already ranks for, and position 27.6 is page three, so it does not survive. SE Ranking keeps the top 50 to 100 keywords, and by clicks this one is at the bottom of the list, so it does not survive. The method that tells you not to pre-filter keeps it right up until its final validation step, which retains only prompts where the domain already appears in the answer, and a page-three page with zero clicks is not being cited by much of anything.

Every published method deletes the single largest piece of first-party demand evidence we have in our own category. That is not a hypothetical failure mode, and we are not describing someone else's data.

05The method

Six steps. The first five are free and take an afternoon.

1. Export a full query report, not a top-queries report. In Search Console, open Performance, set the date range to the last 28 days, and export queries with impressions, clicks, CTR and average position. Twenty-eight days is a deliberate choice rather than a round number: it covers four full weekly cycles, so a B2B pattern that is quiet at weekends is not distorted by where the window happens to land.

2. End the window three days before today. Search Console data lags by roughly two to three days and is revised after it first appears. If you build a set from the last three days of data you are building on numbers that will change underneath you. Cut the window short and re-read the whole thing next time rather than appending.

3. Strip the queries that are not people. A meaningful share of impressions on any site that writes about AI tooling comes from other AI systems grounding an answer, not from humans. They are recognisable: quoted phrases, operator strings, long -site:reddit.com -site:twitter.com exclusion chains, and queries prefixed with stray characters. On our own property these ran to about 31 percent of impressions in the category. Filter them out with a regex excluding quote marks, site: and similar operators. Their zero click-through rate is expected and is not a problem to fix.

4. Sort by impressions and take the top 25. Not by clicks, not by position, and not by your own sense of which queries matter. Twenty-five is enough to carry your buyers' vocabulary and few enough that the long tail of single-impression queries does not drown the signal.

5. Rewrite each one as a question, keeping their words. A search query and a prompt are not the same shape. ai citation tracking becomes something closer to which tool tracks AI citations across engines. The rewrite should keep the buyer's own vocabulary rather than translating it into your product's language, because the phrasing is the part you observed and the rest is your assumption.

6. Prune by consequence. This is the only place your judgement should override the data. Drop any question where losing the answer would not plausibly cost a deal, and any question you could not influence within two quarters. Both tests are borrowed from published guidance and both are good.

06What Search Console cannot tell you

A method is only as trustworthy as the list of things it refuses to claim, and this one has a real list.

It does not report prompt volume. Search Console tells you what people typed into Google. It does not tell you what they typed into ChatGPT, and no engine publishes that for private assistant conversations. Anyone quoting a prompt volume figure for a specific question is estimating from something else. Search queries are the closest observable proxy for how buyers phrase a problem, and a proxy is all they are.

Its AI reporting is thinner than it sounds. Google added Search Generative AI performance reporting in June 2026, covering impressions for AI Overviews, AI Mode and AI features in Discover, broken down by page, country, device and date. It does not report clicks, click-through rate, or the queries behind those surfaces. So Search Console can tell you that AI features showed your page, and cannot tell you what was asked.

It says nothing about assistants that do not use Google. A grounded prompt set built this way inherits Google's demand distribution. That is a genuine limitation and the honest response to it is that no better first-party source exists, not that the limitation does not matter.

It cannot tell you why. Impressions and position describe what happened. Every explanation you attach is a hypothesis, and the useful discipline is to keep observed facts and hypotheses visibly separate in whatever you report internally.

07Why one run proves nothing

Having chosen the right questions, the next mistake is measuring them once. AI answers are probabilistic and they are unstable at a scale that surprises people the first time they measure it.

A study of 530,875 citations found that 40 to 60 percent of the domains cited in AI responses are completely different a month later for identical questions, rising to 70 to 90 percent when comparing across about six months. Daily testing across four engines found 69 percent of a typical answer's sources changed from one day to the next. SE Ranking's own AI Mode research found only 35 percent of domains repeat between runs, with two thirds vanishing.

Two practical consequences follow. First, a screenshot of an answer proves nothing, in either direction: the answer that names your competitor today may name you tomorrow for reasons neither of you caused. Second, any tool that reports a single visibility number without a series behind it is selling you a coin flip with a decimal point. Published guidance converges on running your set for at least 30 days before drawing conclusions, and that is the right order of magnitude.

08How many prompts, and when to prune

Published guidance is unusually consistent here. SE Ranking recommends 20 to 40 in total, split roughly into 10 to 20 awareness prompts, 20 to 30 consideration prompts and 5 to 10 brand evaluation prompts. Other guidance suggests 25 to 40 across two or three engines, run daily for at least 30 days and pruned monthly. A commonly repeated ratio is around 75 percent unbranded to 25 percent branded, on the reasoning that vendor-aware prompts are where displacement shows up first.

Those numbers are fine. The reason they matter less than they appear is that a tracked set is a running cost, not a fixed asset: every prompt you add is measured again on every run, forever, and the discipline that keeps the set useful is removal rather than selection. The two pruning tests worth keeping are the consequence tests from step six. If losing the answer would not plausibly cost a deal, the prompt is a vanity metric. If you could not influence the answer within two quarters, it is a weather report.

One addition, which follows from the argument above: re-ground the set on a schedule rather than treating it as permanent. Buyer vocabulary moves, and a prompt set grounded once in March is an invented prompt set by September. Re-reading the same rolling window and regenerating gives you a set that tracks the market rather than your memory of it.

09Doing this by hand, or not

Everything above is achievable with a Search Console export, a spreadsheet, and an afternoon. For a set of twenty or thirty questions on two engines, checked monthly, that is genuinely the right answer, and it is what we would tell a small team to do before buying any tool in this category.

The manual method breaks in four specific places rather than generally. It breaks when you need history you did not think to record, because the volatility figures above mean a series is the measurement and a series cannot be reconstructed after the fact. It breaks when several people need the same view and the spreadsheet develops opinions. It breaks when the re-grounding step needs to happen on a schedule rather than when somebody remembers. And it breaks when you want the sources each engine cited rather than just whether you appeared, because collecting those by hand across four engines and thirty questions is a day's work per run.

That is the honest boundary. Below it, a spreadsheet wins. Above it, the thing you are buying is durability and comparability, not intelligence.

Frequently asked questions