Guides / Get Found

By DLVX · 12 min read · Updated 20 October 2026 · Part 9 of 10 in Get Found

Measure AI visibility without fooling yourself

  • Measure AI visibility as a rate across many valid answers, never as a position in one screenshot.
  • Read the free first-party reports first: Search Console's AI report and Bing's AI Performance report.
  • Look beyond one run: retain the counts, repeat the same questions and inspect variation. Repeated prompts can be dependent; a simple standard-error threshold is only a screening heuristic.
On this page

What are we actually counting?

Most AI visibility reports fail before the first number. They mix five different things into one score, and then nobody can say what moved or why. So before we look at any dashboard, we agree on the words.

Our competitor says ChatGPT "recommends" them. How do we know where we stand?

We would ask what they actually counted. A business can be named in an answer without being recommended, cited without being named, and recommended without anyone visiting. Those are separate outcomes, and each one is worth a different amount of money.

OutcomeWhat we recordWhat it does not prove
MentionYour name appears in the answer text.That the answer approves of you.
CitationA source link points to one of your pages.That you are on the shortlist. Engines cite an explainer and recommend someone else all the time.
RecommendationThe answer proposes you as a fit for the buyer.That anyone clicked.
VisitA session arrives from an AI surface.That the visitor was a buyer.
LeadAn enquiry that meets your fit criteria.That it became revenue.
MentionCitationRecommendedVisitLeadCount each one on its own line.
Five different outcomes. A report that adds them into one score hides which one moved.

Our rule is simple. Each outcome gets its own line, its own denominator and its own date. A mention rate of 30 percent means "named in 30 of every 100 valid answers to our fixed buyer questions", never "30 percent of the market". If a supplier cannot tell you the denominator, the number is decoration.

Start with the data the engines give you

Before anyone builds a prompt set, read what Google and Microsoft already report for free. These numbers come from the engines themselves, so they beat any scraper on accuracy. Their weakness is that they are thin.

SourceWhat it showsWhat it cannot show
Google Search Console, generative AI reportImpressions in AI Overviews and AI Mode by page, country, device and date. Announced 3 June 2026.Clicks. No query text. It counts by property, so two of your results in one feature are one impression.
Google Search Console, normal Performance reportAI Overview and AI Mode traffic is included under Web search, mixed with everything else.Which clicks came from an AI feature.
Bing Webmaster Tools, AI PerformanceTotal citations, cited pages and a sample of grounding queries for Copilot. Public preview since February 2026, per a detailed secondary review.Clicks, rank, and there is no API yet.
Your analytics, AI channelSessions that arrive with a recognisable AI referrer.Visits that arrive with no referrer. Treat this channel as a floor, never the total.
Changed in 2026: you can switch AI features off

Announced in June 2026 and available to all sites since 31 August 2026, Search Console has a setting that excludes a site from AI Overviews and AI Mode. An excluded site gets no AI impressions or traffic from those features, and the generative AI report can go missing. The setting does not stop training, which is a separate control, and it does not remove you from Search. Check it before you diagnose a zero.

We read these monthly for every site we run. They answer two questions cheaply: are we appearing in AI features at all, and which pages carry it. Everything after this section is about the questions they cannot answer, mostly which answers name you and whether that is changing.

Why one screenshot proves nothing

The single most common mistake we see is a screenshot. Someone asks ChatGPT a question, sees their brand, and the screenshot goes into a board deck. A week later a colleague asks the same question and gets a different list.

Under 1 in 100chance that ChatGPT or Google AI returned the same brand list twice for the same prompt (SparkToro, data November to December 2025, published January 2026)
21 to 58%overlap between two runs of the same question in Google AI Mode, by category, in our own Bangkok study, October 2026
9.2%average URL overlap across three same-day AI Mode runs (SE Ranking, August 2025)

SparkToro's study, published 28 January 2026, had 600 volunteers run 12 prompts through ChatGPT, Claude and Google AI, close to 3,000 runs. The same list came back less than once in a hundred times, and the same list in the same order about once in a thousand. Their useful conclusion: the share of runs that name a brand looked reasonably consistent across many runs, while rank position did not. Worth knowing: the study is not peer reviewed and a co-author works for a company that sells AI tracking.

Our own numbers point the same way. In our study of who AI recommends in Bangkok, published 3 October 2026, we asked Google AI Mode 80 buyer questions three times each. Two runs of the same question shared between 21 percent (hotels) and 58 percent (international schools) of the businesses named.

Sources drift too. SISTRIX tracked 82,619 prompts for 17 weeks and read weekly churn of new sources at about 5 percent for AI Overviews against about 56 percent for AI Mode and 74 percent for ChatGPT search (values read from their chart). Profound, a tracking vendor, measured month to month citation drift between 40 and 59 percent across four engines in mid 2025. SE Ranking found AI Mode shared only 9.2 percent of URLs across three runs on the same day.

So is AI visibility even measurable?

Yes, as a rate, not as a position. A rate over many valid answers is a reasonable signal. A position in one answer is noise. That single shift, from "where do we rank in ChatGPT" to "in what share of answers are we named", is most of the method.

Build a prompt set you will not regret

The prompt set is the ruler. If you change it every month, you have no ruler. These are the rules we run on every site.

  1. Start with ten unbranded buyer questions, grow to 15 to 50 for measurement. Phrase them the way a buyer types, in each language you sell in. Take them from real enquiries, sales calls and long questions in Search Console. Never include your brand name: a question that names you will name you back, and the rate becomes meaningless.
  2. Freeze the file. Keep it in one place with a date. Add questions when the business adds a service. Never quietly rewrite the ones that perform badly.
  3. Run logged out, in a fresh context. Personalised chats carry history. In our own harvests a logged-in chat put the brand name into the hidden search; a logged-out temporary chat did not.
  4. Record the conditions. Date, engine, mode, location, language, signed-in state. Different conditions are a different population, not a trend.
  5. Throw out what is not an answer. A quota message, a sign-in wall, a bot check or a cut-off reply is not an answer. We mark every record valid or invalid with a reason, and invalid records never enter a count.
  6. Code the answer three ways. Named, cited (with the URL), recommended. Plus anything wrong: an old address, a service you dropped, a confused identity.

The prompts are test inputs, not search volume. They tell you how an engine answers a fixed set of buyer questions. They do not tell you how many people ask them. Say that in every report.

The maths that stops false wins

This is the part most reports skip, and it is the reason most "wins" are not wins. Answers vary run to run, so a rate always comes with noise. Our engine applies four rules before anything is called a result.

  • Pool four weekly runs into one figure, and only score it once there are at least 60 valid answers.
  • Never quote a rate built on fewer than 10 valid answers. Not in a report, not in a meeting.
  • Use two standard errors only as an initial screening heuristic under an independence assumption. Repeated questions, shared model versions and overlapping source retrieval can violate that assumption.
  • If the window spans a Google core or spam update, the result is inconclusive unless the effect clears twice the usual noise band.

Show me with real numbers. When is a jump a jump?

Say the four weeks before a change gave 18 mentions in 72 valid answers. That is 25 percent. The four weeks after gave 27 in 75, which is 36 percent. Eleven points up. It feels like a win.

Before: p1 = 18 / 72 = 0.25
After:  p2 = 27 / 75 = 0.36

Variance before = 0.25 x 0.75 / 72 = 0.0026
Variance after  = 0.36 x 0.64 / 75 = 0.0031
Standard error of the difference = square root of (0.0026 + 0.0031) = 0.075
Two standard errors = 0.15, so 15 points

Change = 11 points. Smaller than 15. Not a win yet.

Now say the after period gave 40 mentions in 80 answers, 50 percent. The standard error barely moves (about 7.6 points, so a threshold near 15), and the change is 25 points. That is a stronger observed change under the simplified calculation. Report both counts and windows, inspect prompt-level dependence, and do not attribute the change to your work without a suitable comparison.

The update rule matters more than it sounds. Google ran a spam update starting 24 September 2026 that lasted 13 days and 16 hours, after two core updates earlier in the year. Any before and after that straddles one of those is measuring Google, not your page.

What this means for small sites

With 20 prompts and one run a week, you need about three weeks just to reach 60 valid answers, and the four week pool sets the pace anyway. That is fine. A slow number beats a fast flattering one, and it stops a business from rewriting a page that was working.

What a mention is worth in clicks

Visibility only matters if it reaches the business. The research here is loud and it disagrees, so we read it as a range, not a verdict.

StudyWhat it measuredWhat we take from it
Pew Research Center, July 2025900 US adults' browsing in March 2025: 8% of visits with an AI summary clicked a normal result, against 15% without one. 1% clicked a link inside the summary.Summaries take clicks. Links inside them get few.
Ahrefs, February 2026300,000 keywords, desktop: position 1 click rate about 58% below the forecast without an Overview.The top spot is worth less on informational queries. Correlation, vendor data.
Seer Interactive, November 2025Organic click rate on AI Overview queries fell from 1.76% to 0.61%. Brands cited in the overview got 0.70% against 0.52% uncited.Being cited softens the loss. Seer notes it cannot prove the citation causes it.
SparkToro, June 202668.01% of US Google searches ended without a click in January to April 2026. Only 0.34% reached AI Mode.Zero click is the norm. The panels changed across years, so the trend is approximate.

Then there is AI referral traffic itself. SE Ranking's panel of 101,574 sites (June 2026) puts AI engines at 0.32 percent of all website traffic in 2026, with Google organic sending about 134 times more. Similarweb (July 2026) shows the AI side shifting fast: ChatGPT's share of generative AI web visits fell from about 76 to about 53 percent between June 2025 and May 2026, while Gemini rose to about 27 percent, a vendor estimate with little method published.

On conversion, the studies flatly disagree. Ahrefs reported AI visitors converting about 23 times better than organic on its own site, and warned against using that as a benchmark. Amsive compared 54 sites and found no significant difference (4.87 against 4.60 percent, p = 0.794). Semrush estimated an AI search visitor as worth about 4.4 times an organic one, a vendor estimate with little method published.

So should we chase AI traffic or not?

Chase being the right answer for your buyer questions, and keep Google organic healthy, because that is still where most of your visits come from. Then measure your own conversion from both channels. Your own numbers beat any of these studies for your business.

Mistakes we see, and the checklist we use

Mistakes

  • Testing only your own brand name, then claiming buyers discover you.
  • Rewriting the prompt set when the numbers disappoint.
  • Counting a quota page or a sign-in wall as "not mentioned".
  • Reporting a rate from eight answers.
  • Calling an 11 point rise a win when the noise is 15.
  • Reading the analytics AI channel as the full total. Many AI visits arrive with no referrer.
  • Comparing a month that contains a Google update with one that does not, and crediting your page.

Checklist

  1. Definitions agreed. Mention, citation, recommendation, visit, lead.
  2. First-party read. Search Console AI report, Bing AI Performance, AI control setting checked.
  3. Prompt file frozen. 15 to 50 unbranded buyer questions, dated, per language.
  4. Conditions logged. Engine, mode, location, logged out, fresh context.
  5. Validity coded. Every record valid or invalid, with a reason.
  6. Pooled and gated. Repeat runs and preserve denominators; inspect dependence before treating a standard-error calculation as evidence of a change.
  7. Confounders checked. Google updates on the Search Status Dashboard.
  8. Business line kept. Visits and qualified leads reported next to visibility, never merged into it.

Next in the track: how we triage any SEO tactic in ten minutes. If you want the inputs that move these numbers, read query fan-out and pages that AI answers can cite.

Sources and dates

Every external fact in this guide was read on 20 October 2026. Each source shows its publisher and the date it was published or last updated, and we flag what changed in 2026. Studies are their authors' measurements, usually correlations, not promises of results. Our own numbers come from our own Search Console and tools and say so where they appear.

Want this run on your own site? Start with the AI visibility diagnostic, or see the whole Get Found track.