By Mark Buraga, Independent SEO Consultant at Growth Engine PH
Last updated: 5 October 2026
Same 15 questions. Same day. Google AI Mode returned 281 cited sources to a Philippines session and 57 to a United States session. Every AI visibility score you have ever been shown came out of one session like that, and nobody told you which one.
That is the real difficulty in how to measure AI visibility today. The answer layer is non-deterministic. Put the same question to the same engine twice and it assembles a different set of sources, so when a vendor calls their number accurate, no test you could run would prove them wrong. Accuracy is the wrong property to shop for, because an accuracy claim against a moving system cannot fail.
Reproducibility is the property you can check. This page is our method in full, published with the numbers that make us look worse rather than the ones that would sell better.
What does it actually mean to measure AI visibility?
Measuring AI visibility means running a fixed set of buyer questions through an answer engine on a fixed schedule and recording two separate things per question: whether your site appeared as a cited source, and whether your brand was named inside the answer text.
Those two are not one blurred score. A citation is the source link the engine attributes part of its answer to, earned structurally by being retrievable, which our retrievability diagnostic covers. A mention is your brand named inside the answer text, earned off the page through market consensus. They move independently, so each gap points at different work.
Want a one-off check on your own site instead of an ongoing standard? That is a different job, and here is the 20-minute version: test it yourself.
Why accuracy is the wrong thing to ask an AI visibility tool for
No AI visibility tool can be accurate, because the same prompt put to the same engine on the same day returns a different set of sources, so the only property a buyer can actually verify is whether the method is reproducible.
The instability is measured, not asserted. According to research reported by Kevin Indig at Search Engine Land on 10 June 2026, within-model variance from sampling alone runs 10% to 34% on identical prompts; only 2.2% of citations survive three runs of the same ChatGPT prompt, across 815,000 prompt-page pairs; and Google AI Mode replaces 56% of its cited sources week to week while ChatGPT replaces 74%, across 82,619 prompts tracked for 17 weeks. We have not independently verified those datasets.
Our own corpus reproduces that instability. On a frozen 15-prompt set captured on one script for seven consecutive weeks, total cited sources ranged from 57 to 243 with nothing changed on our side. That series, its noise floor and what it does to a monthly score are taken apart in our volatility post; this page stays on the procedure. Most reporting cannot see this surface at all, which is why dashboards miss this while every line on them stays green.
The four questions that separate a method from a score
A reproducible AI visibility method answers four questions in public: which prompts, captured how, what is recorded, and what is deliberately not measured.
That is the whole audit, and it works on any vendor. A method answers all four on a page you can read before signing. A score answers none and asks for trust instead.
Which prompts make the 15, and how are they chosen?
Growth Engine PH tracks exactly 15 prompts per engagement through Google AI Mode, split 5 brand-direct, 5 service-intent and 5 competitor-comparison, chosen once at onboarding from the client’s own buyer language and never changed mid-engagement.
Fifteen is a working compromise, not a discovered constant. In practice, 15 prompts is what one person can capture by hand on the same weekday inside about 30 minutes, without the run turning into a chore and getting skipped. The three-way split matters more than the total. A single blended figure hides the gap between an engine knowing a brand and an engine recommending a brand to a buyer, and the two results come apart badly. We measured the gap across eight captures in 2026: Growth Engine PH was cited on 39 of 40 brand-direct captures (98%) and on 5 of 80 buying-intent captures (6%). Any set built by Growth Engine PH keeps the buckets apart: a mixed score can rise while the half that wins clients stays flat.
The three buckets
The three buckets, and the shape of prompt each one holds:
| Bucket | Count | What it tests | Example shape |
|---|---|---|---|
| Brand-direct | 5 | Whether the engine knows the entity | legitimacy, reviews, founder, pricing, location |
| Service-intent | 5 | Whether it cites you on commercial questions | “best [service] for [persona]”, “[service] cost” |
| Competitor-comparison | 5 | Whether it names you when ranking providers | “top [service] firms in [region]”, “[brand] alternatives” |
The four selection rules
Four rules govern which questions make the set, and all four are applied once, at onboarding, in front of the client.
First, at least 3 of the 15 (20%) must carry a city, country or region, because local intent is where Google AI Mode most often reaches for cited sources. Second, exactly 2 must be questions a competitor would expect to win, so the set holds a real test and not only a comfortable one. Third, no more than 1 may be a pure brand-volume query, because the point is discovery rather than brand defence. Finally, every prompt uses the buyer’s vocabulary rather than ours. In practice the fourth rule bites hardest on our own set, frozen in June 2026 and run through Google AI Mode since: the set tests the word “agency” even though Growth Engine PH is a consulting practice and never uses the word, because “agency” is what buyers type.
Why the set is frozen for the life of the engagement
The prompt set is frozen because the point of the measurement is the week-over-week delta, and changing a prompt mid-engagement destroys every comparison that came before it.
Businesses change, so eventually the prompts do too, under a rule: a set that needs changing gets version-bumped into a new series rather than overwritten. What is never allowed is quietly swapping a prompt that has been performing badly. That single move is how an unpublished method launders a bad quarter, invisibly, unless the set was published first.
Why do the capture conditions belong in the method?
Every capture runs against one engine surface, Google AI Mode, in a clean session with the geography set to the client’s primary market, with each prompt typed verbatim and never refined, because changing any one of those conditions changes the result more than anything the client did that week.
The conditions are four lines, applied in order. First, a clean or signed-out session, so no earlier query shapes the answer. Second, geography set explicitly rather than inherited from the network. Third, the prompt typed verbatim, with no autocomplete and no refining follow-up. Finally, the full answer and the source panel captured before anything is counted. In practice the four lines take under a minute per prompt and remove the largest known sources of drift. We measured what skipping the second line costs on 13 June 2026: the same 15 prompts returned 281 cited sources in a Philippines session and 57 in a United States session, an 80% fall, with Growth Engine PH cited on 6 answers in one and 3 in the other. A capture record without the four conditions cannot be compared with the capture before it.
The two sessions of 13 June 2026
On 13 June 2026 we ran the first real capture of Growth Engine PH’s own AI visibility, and the first attempt never got past the wall: the capture script was bounced to Google’s CAPTCHA page and returned nothing. A persistent browser profile cleared the block. Google AI Mode then returned 281 cited sources across the 15 answers, with our domain cited on six and our brand named in six.
The same 15 prompts were run again that day with the session geography forced to the United States. 57 cited sources instead of 281. Our citations fell from six to three. The mentions held at six. That United States run sits outside the weekly series, which is captured in the Philippines throughout.
Nothing about the business had changed in the intervening minutes. One caveat belongs with that number, and it is the kind an unpublished method never has to make: those two runs were not captured on identical harnesses. The Philippines baseline ran through a headed browser context after the CAPTCHA bounce, the United States run through the batch runner we standardised on later. Geography is the largest visible difference and not the only one, and the confound is visible only because the conditions were written down. An AI visibility figure without its capture conditions is an anecdote.
What gets recorded for every single prompt?
Each of the 15 prompts produces the same six recorded fields every week: the number of cited sources, whether the client’s own domain was among the sources, which exact URLs, whether the brand was named in the answer text, the context of any mention, and which competitors were named instead.
Two more items are captured beyond the six fields. First, every source domain in the panel, which is what makes cross-week aggregation possible. Second, a full screenshot of the answer and the source panel, kept as evidence rather than data. In practice the structured record is what every report is built from, and the screenshot settles disputes about what Google AI Mode showed. We measured how much weight the domain field carries in 2026: across the 10 buying-intent prompts alone, 568 distinct domains appeared over eight weeks, and Growth Engine PH was among the cited domains in 6% of those 80 captures. The Growth Engine PH audit standard puts the rule flatly: “The audit counts are the truth for AI visibility.”
The six fields
The six fields, and what each one holds:
| Field | What it holds |
|---|---|
| Citation count | Total sources in that answer’s panel |
| Client cited | True or false for the client’s domain |
| Client citation URLs | The exact pages cited, not the domain |
| Mention present | True or false, inside the answer body |
| Mention context | Positive, neutral, comparison, competitor-adjacent |
| Competitor mentions | Every rival brand named in that answer |
Eight weeks of AI visibility measurement on our own site, in two numbers
Across eight consecutive weekly audits of our own site, Growth Engine PH was cited on 39 of 40 brand-direct prompt captures and on 5 of 80 buying-intent captures, which is a near-total split between being findable when someone already knows our name and being invisible when they do not.
That second number is the one a case study would have left out: on the questions that would win us a client, we are absent 94% of the time. The weekly totals behind those two numbers moved a great deal on their own, from 57 to 243 cited sources on an unchanged script, and that noise is a separate finding with its own post, linked below.
Which domains win the answers we lose
Underneath those totals, 568 distinct domains appeared across just the 10 buying-intent prompts. Reddit led with 34 appearances on 8 of the 10. LinkedIn was second with 24, also on 8 of 10. Neither is a site anyone can optimise.
Read that as a current-state observation rather than a law. Reddit’s own chief executive has questioned what the arrangement returns, telling Ars Technica on 7 August 2026 that “we’re still looking for that win-win”. Nor is the doom framing settled: Shopify reported that AI search is driving more traffic and sales rather than replacing Google (TechCrunch, 6 August 2026). Three companion posts take each finding apart properly: the brand split, the Reddit and LinkedIn count, and the week-to-week noise.
What does this AI visibility method not measure?
This method measures visibility on one engine, in one geography, on a 15-prompt sample, with a human in the loop, and it never measures revenue.
Fifteen prompts is a sample, not a census. One prompt moving takes the headline number with it by nearly 7%, and no amount of weekly discipline fixes that.
One engine. We capture Google AI Mode and nothing else, so this method says nothing about what ChatGPT, Perplexity, Claude or any other answer engine is doing with your brand. Those surfaces also shift unpredictably, and why this shifts matters more than any single week’s reading.
One geography per run. The paired runs above show how much of a result can belong to the session rather than the brand. And a human verifies every count, which introduces judgement and caps the volume at what one person can check.
It measures visibility, never attribution. Citations are not conversions. Whether any of this reached a buyer sits with the client’s own CRM and UTM hygiene.
The upgrade path is fixed in that order: repetitions per prompt first, then additional engines.
Where do thoughtful people disagree with this method?
Two credible published positions argue against a frozen 15-prompt weekly capture: one says the sample is too small and should be roughly 40 prompts run five times each, and one says prompt sampling is not credible at all and first-party grounding data should replace it.
Kevin Indig says the sample is underpowered. His published framework recommends around 40 seed prompts for a business-to-business example, split 12 brand, 12 category and 16 problem-focused, with five repetitions per prompt per platform per week, tracked separately by platform with confidence intervals and persona variants (Search Engine Land, 10 June 2026). He opens by conceding the critics have a case, writing that “the prompt-tracking backlash is only half-right”. Against his standard, one weekly pass over 15 prompts is a small sample, and he is right about that. We concede it in the limits above, in those terms.
What the 15-prompt set buys instead is direction on a fixed instrument at a cost that survives a busy month. Our standard caps the audit at about 30 minutes per client per week, and treats that cap as non-negotiable. A 40-prompt, five-repetition design is a 200-run job per client per week, which in a solo practice means the audit gets skipped when the month gets heavy, and the skipped weeks are exactly the weeks something changed. A method that always runs beats a better-powered one that runs when there is time.
Christopher Penn says prompt sampling is not credible at all. His position is that nobody can honestly say how a brand ranks inside a large language model, because mixture-of-experts routing sends the same query down different paths, and that the real measurement stack is customers telling you how they found you, then grounding data from Search Console and Bing Webmaster Tools, then CDN bot logs piped into analytics (Trust Insights, July 2026).
We agree on the ceiling and disagree on the substitute. Grounding and impression data tells you how often a surface appeared. It never tells you what the answer said, whether the brand was named, or which competitors were named instead, and those are the reads this method exists to produce. So we run both where the data exists. The Search Console generative-search report joins the weekly audit on any property where Google has switched it on, with three limits stated every time: impressions only, Google surfaces only, and an uneven per-property rollout, so its absence is not evidence of no visibility. At our last check it had not reached our own property.
| Design | Prompts and repetitions | Strongest at | Gives up |
|---|---|---|---|
| Growth Engine PH, weekly frozen set | 15 prompts, 1 pass, weekly | Direction on a constant instrument | Power on any single reading |
| Kevin Indig’s framework | ~40 prompts, 5 repetitions | Confidence intervals, per-platform splits | Weekly feasibility at boutique cost |
| Christopher Penn’s stack | No prompt sampling | Census-level first-party grounding | What the answer said, who was named |
Two published methods that disagree is a functioning field. Two unpublished scores that disagree is a coin toss.
How do you hold any vendor, including us, to this standard?
Before you accept an AI visibility number from anyone, ask for the prompt set, the capture conditions, the recorded fields and the limits, and treat a refusal on any of the four as the answer.
A good answer is boring and specific: here are the prompts, the engine and geography, the fields, and what we do not claim. An evasive answer talks about proprietary models or a scoring algorithm and never puts a prompt on the page.
This audit is not an upsell. It runs weekly as a standard deliverable on every Growth Engine PH retainer tier, with a baseline captured in week one before any work begins, so the “before” exists whether or not the work goes well.
Get your own baseline read. Fifteen prompts, your buyer language, captured before any work starts, and you keep the record either way: baseline read.
A number you cannot reproduce is a number you cannot audit. Ours is on this page, including the parts that make us look worse.
Frequently Asked Questions
My page ranks top 5 but is not cited in the AI answer. What do I check first?
Check the capture conditions before you check the page. Run the prompt three times in a clean session with the geography explicitly set, because our own paired runs returned 281 cited sources in one session and 57 in another on the same day. If the page is genuinely absent across repeated clean runs, the diagnostic moves to retrievability, and the decoupling between rank and citation is where that trail starts.
What does each AI engine want before it cites you?
We can only answer this for one engine honestly. Our method captures Google AI Mode, so we describe what we observe there and nothing about ChatGPT, Perplexity or Claude. On AI Mode, what gets cited is retrievable structure: a page whose entity is confirmable and whose answer lifts out as a self-contained passage. Anything we said about other engines would be inference dressed as measurement.
What happens after you get cited?
Often nothing you can see, and that is why this method separates visibility from attribution. Citations are the measurement; conversion sits with your CRM and UTM hygiene. If citations rise and nothing else moves, the usual cause is that the answer is not naming your brand, which is the mention side rather than the citation side.
How many prompts do I need for an AI citation tracking method?
Enough to be stable, few enough to run every week without exception. We use 15 and publish that one prompt moves the headline by nearly 7%; a credible published alternative recommends around 40 prompts run five times each. Pick the largest set you will genuinely run weekly, then freeze it.
Should I measure AI visibility weekly or monthly?
Weekly. The week-to-week noise on an unchanged prompt set is large enough that a monthly snapshot samples it once and reports the sample as a trend. Our own measured noise floor, and how to read a score against it, is the subject of our volatility post. Capture weekly, compare rolling readings, and never treat one week’s number as the level.
Why keep the same prompts instead of updating them?
Because the whole value is the week-over-week delta, and a changed prompt breaks every comparison behind it. When a set genuinely needs changing, we version it and start a new series rather than overwriting the old one.