An AuditSpark.io whitepaper on probing real category questions across engines, and reading what comes back.
Website intelligence that sparks action.
TL;DR
- Readiness and visibility are different measurements. A site can be perfectly accessible, clearly structured, and well evidenced, and still never be mentioned when a buyer asks an engine for a recommendation.
- You cannot measure this from Google Search Console. Google reports AI Overviews and AI Mode traffic inside Web Search totals, and it tells you nothing about how ChatGPT, Claude, Perplexity, or Gemini describe your brand.
- Engines disagree with each other, substantially. Academic work finds that platforms differ in both how many sources they cite and how much influence a cited page has on the final answer. Third-party industry studies report very low source overlap between engines. Measuring one engine tells you about one engine.
- Absence is not the only risk. Engines can describe your pricing, your services, or your specialty inaccurately, and say it with confidence. Misrepresentation is often more damaging than silence.
- One snapshot goes stale. Research on GEO explicitly warns that strategies tuned to fixed engine behavior overfit and degrade when engines update. That makes visibility a monitoring problem, not a one-time audit.
- The measurable quantities are concrete: mention rate, citation rate, share of voice against competitors, position, sentiment, factual accuracy, recommendation strength, and cross-engine variance.
Start here: subscribe to AuditSpark.io to run AI Visibility checks across engines and see whether you are being mentioned, cited, recommended, or misrepresented.
Executive summary
The first three papers in this series were about readiness. Can engines reach your site, understand it, trust it, and find something worth quoting. This paper is about the question those three cannot answer: when a real buyer asks a real question, does the engine actually name you?
That gap is wider than most teams expect. Readiness is a property of your website, so you can inspect it directly. Visibility is a property of the engines, and it lives inside black boxes you do not control. You can only learn it by asking, which means active probing rather than passive analytics. A site can score well on every readiness layer and still be absent from the answers that matter, because the engine weighed other sources, or knows a competitor better, or simply never retrieved you.
The measurement infrastructure most teams rely on does not cover this. Google reports AI Overviews and AI Mode traffic within its Web Search totals, so it gives you no distinct view even of Google's own AI surfaces, and by definition nothing at all about ChatGPT, Claude, Perplexity, or Gemini. There is no console for the answer layer. If you want to know what engines say about you, someone has to ask them and write down what comes back.
The case for asking more than one engine is now empirical rather than intuitive. A 2026 measurement study of over twenty thousand citations across ChatGPT, Google's AI surfaces, and Perplexity found that citation breadth and citation depth diverge, meaning some platforms cite many sources while others cite few but lean on them far more heavily. Third-party industry research goes further, reporting very low overlap in which sources different engines cite for the same query. The practical consequence is blunt: a result from one engine is evidence about that engine, not about your visibility in general.
The case for measuring repeatedly is also grounded. Research on GEO explicitly identifies that methods tuned to fixed engine behavior tend to overfit and degrade when engines update, treating generative engines as moving targets rather than stable systems. A visibility snapshot is therefore perishable. It tells you where you stood on the day you asked.
This paper defines Layer 6 of the AuditSpark.io framework. It covers what to measure, how to build a probe set that reflects real buyer questions, how to read the results without fooling yourself, what agencies can sell from it, and a ninety-day plan to establish a baseline and start tracking. It also states clearly what this measurement cannot do, because nothing here guarantees that any change will make an engine recommend you.
Why it matters now
There is no console for the answer layer. Every other channel a marketer manages has a dashboard. Search has Search Console. Ads have a platform. Social has native analytics. The answer layer has nothing. Google's own reporting folds AI Overviews and AI Mode into Web Search, and no engine offers a brand-visibility report. This is the largest unmeasured surface in modern marketing, and it is growing.
Absence leaves no trace. When an engine does not mention you, there is no impression, no click, and no error. Nothing in your analytics distinguishes "we were never considered" from "the market is quiet this month." The only way to detect it is to ask the question yourself and observe the absence.
Engines disagree, so a single check is misleading. If the engines broadly agreed, checking one would be a reasonable proxy. They do not. The research shows meaningful divergence in citation behavior across platforms, and industry studies report that most sources are cited by only one engine. A brand can be well represented on one platform and invisible on another, and a single-engine check would report either as the whole truth.
The target keeps moving. Engines update. Retrieval behavior changes. The GEO literature explicitly warns that strategies fitted to today's engine behavior degrade as engines evolve. That converts visibility from an audit into a monitoring discipline, which is exactly the kind of work that sustains a retainer.
Technical validation
This section separates what is established in peer-reviewable research from what comes from third-party industry studies. Both are useful. They should not be presented with equal weight.
Established: the reporting gap is real
Google's documentation confirms that traffic from AI Overviews and AI Mode is reported within Web Search in Search Console rather than broken out separately. There is no cross-engine equivalent at all. No standard console reports whether ChatGPT named your brand, whether Perplexity cited your page, or whether Claude recommended a competitor. The gap is structural, not a tooling oversight, and it is the reason active probing exists as a practice.
Established: engines diverge in how they cite
A 2026 measurement study proposed a two-stage framework distinguishing citation selection, whether a platform triggers a search and chooses your source at all, from citation absorption, how much a cited page actually contributes language, evidence, or factual support to the final answer. Analyzing 602 controlled prompts across ChatGPT, Google's AI Overview and Gemini, and Perplexity, and over twenty-one thousand search-layer citations, it found that citation breadth and citation depth diverge. Perplexity and Google cited more sources on average, while ChatGPT cited fewer sources but showed substantially higher average citation influence among the pages it did fetch.
The strategic implication is important and easy to miss. Being cited is not one thing. You can be one of eleven sources that barely shaped the answer, or one of three that largely wrote it. A measurement program that counts citations without weighing influence will misread its own results.
Established: fixed tactics decay as engines change
Recent GEO research is explicit that existing methods relying on static heuristics or engine-specific rule distillation are prone to overfitting, and that they degrade when the engine updates. The stated framing is that generative engines are black boxes with changing behavior, and that robustness comes from improving intrinsic content quality rather than fitting to a snapshot of engine preferences. For a measurement program, this is the argument for cadence: what you learn today describes today.
Third-party industry findings, directional
Vendor and industry studies of AI citations report very low overlap between engines. Commonly reported figures include a small single-digit to low-double-digit percentage of shared sources between two major engines, a large majority of sources cited by only one engine, only a small fraction cited by all major engines, and wide differences in how many sources each engine lists per answer. Some studies also report very large variance in citation volume for the same brand across platforms.
These figures are methodology-dependent, come from commercial vendors with an interest in the conclusion, and vary between studies. Cite them as third-party industry research and as directional support for a point the academic work already establishes, which is that engines differ. Do not build a claim that rests on a specific vendor number.
What this measurement cannot tell you
Probing tells you what an engine said, at a moment, for a prompt. It does not tell you why. It does not establish that a change you made caused a change in the result, because you do not control the engine and cannot hold it constant. It cannot guarantee that any intervention will improve your standing. Treat AI visibility measurement the way a good analyst treats survey data: it is real evidence about a population you cannot otherwise observe, and it is noisy, and both of those things are true at once.
Business impact
You cannot manage what nobody is measuring. Every competitor in your category is either being recommended by engines or not, and almost none of them know which. The first business in a category to measure this is operating with information the others do not have, and the cost of acquiring it is low.
Misrepresentation is the underrated risk. Most conversations about AI visibility focus on absence. In practice, the more damaging failure is often a confident, wrong answer: outdated pricing, a service you no longer offer, a location you never served, an attribute that belongs to a competitor. A buyer who receives a wrong answer about you does not know it is wrong. They simply act on it. Factual accuracy is a measurable dimension of a visibility program and it is frequently where the first real finding surfaces.
Share of voice makes the abstract concrete. "Are we visible in AI" is hard for an executive to act on. "In the twenty questions our buyers actually ask, we were named four times and our closest competitor was named fourteen" is not. Competitive share of voice is the single most legible output of this work and usually the one that unlocks budget.
Monitoring is a retainer, and audits are not. A readiness audit is a project. It ends. Visibility measurement, because the engines move and competitors act, is inherently recurring. For an agency, this is the difference between a one-time fee and a line item that renews, and it is justified by the research rather than manufactured.
The practical audit framework
This is Layer 6 of the AuditSpark.io framework, and it works differently from the others. Layers 1 through 5 inspect your website. Layer 6 interrogates the engines. The method is straightforward, and the rigor is in the discipline.
Build a probe set that reflects real buyer questions
Your probe set is the heart of the program, and the most common failure is writing prompts you wish buyers asked rather than the ones they do. Cover four intents:
- Discovery. Open category questions with no brand named. "What is the best tool for X." This is where you find out whether you exist at all.
- Comparative. Head-to-head or shortlist questions. "X versus Y" or "alternatives to Y."
- Recommendation. Explicit advice-seeking with constraints. "What should a small agency use for X on a limited budget."
- Factual and branded. Direct questions about you. "What does [brand] do." "How much does [brand] cost." These are how you detect misrepresentation.
Unbranded discovery and comparative prompts measure whether you are in the consideration set. Branded factual prompts measure whether the engine describes you correctly. You need both, and they answer different questions.
Probe across engines, and record everything
Run the probe set across the engines your buyers actually use, not just the one you use. Where an engine offers both a native mode drawing on internal knowledge and a live web search mode, understand that these can behave very differently, and record which one produced the result. Capture the raw response, not just a score, because the verbatim text is your evidence and your source of qualitative insight.
Score the dimensions that matter
| Metric | What it tells you |
|---|---|
| Mention rate | How often you appear at all across the probe set |
| Citation rate | How often your site is actually cited as a source |
| Position | Where you appear when named, first or fifth |
| Share of voice | Your mentions relative to competitors in the same answers |
| Sentiment | Whether you are described favorably, neutrally, or poorly |
| Factual accuracy | Whether what the engine says about you is true |
| Recommendation strength | Whether you are merely mentioned or actively recommended |
| Cross-engine variance | How much the picture changes between engines |
Recommendation strength deserves its own line, because being mentioned and being recommended are not the same outcome and a program that conflates them will report false comfort.
Read the results honestly
Three disciplines separate a useful program from a vanity one. First, treat a single run as noisy. Engines are non-deterministic, and one answer is an anecdote. Second, do not infer causation from a change, because you did not control the engine. Third, hold the probe set stable over time if you want to compare periods, since changing the questions changes the results independently of anything you did. If you must add prompts, track the old set alongside so the trend line stays honest.
The agency and freelancer opportunity
This is the layer that turns GEO from a project into a practice.
The offer is easy to explain and hard to argue with. You are going to ask the engines the questions your client's buyers actually ask, record what comes back, and report whether the client is being mentioned, cited, recommended, or described incorrectly, and which competitors are being named instead. No promises are required, and none should be made. The deliverable is evidence.
The first report is almost always the most valuable meeting you will have with that client, because it converts an abstract anxiety into a specific, competitive, and often uncomfortable fact. A share-of-voice table showing that a competitor is named three times more often than the client is a stronger argument for a content and trust engagement than any pitch deck.
From there the work compounds naturally. The visibility report identifies where you are weak, which points back to the readiness layers in the previous papers, which is billable improvement work. Then the next measurement cycle tells you whether the picture changed, which justifies the next cycle. Because engines update and competitors move, there is no natural end point, which is what makes this a retainer rather than a project. Price it as recurring monitoring with a quarterly strategic review, and be candid that you are measuring, not guaranteeing. Clients respect that framing, and it protects you.
The next and final paper in this series covers how to package, price, and sell the full ladder from free readiness check to recurring visibility monitoring.
Common mistakes
Measuring one engine and generalizing. The research shows engines diverge in what they cite and how heavily they lean on it. A single-engine result is evidence about that engine only.
Measuring once. Engines update, and the literature warns explicitly that behavior fitted to a snapshot degrades. A one-time visibility check has a short shelf life. Establish a cadence or do not bother.
Only asking branded questions. Asking an engine "what does my company do" will usually produce something. It tells you nothing about whether you enter the consideration set when a buyer asks an unbranded category question, which is where the commercial value sits.
Confusing mention with recommendation. Being listed among eight options is not being recommended. Track them separately or you will report progress you have not made.
Treating a citation as a citation. The measurement research distinguishes selection from absorption. Being one of eleven barely-used sources is not the same as being the page that shaped the answer.
Reading a single run as a signal. These systems are non-deterministic. One answer is an anecdote, not a measurement.
Claiming causation. You changed your page, the engine changed its answer, and you cannot prove those are connected. Report the correlation, keep the humility, and keep measuring.
Changing the probe set and comparing anyway. If the questions change, the trend line is meaningless. Hold the set stable, and version it deliberately when you must evolve it.
The 30, 60, 90 day action plan
Days 1 to 30, establish an honest baseline. Build a probe set of fifteen to twenty-five questions covering discovery, comparative, recommendation, and branded factual intents, written to reflect what buyers actually ask rather than what you wish they asked. Identify the three to five competitors you expect to see. Run the set across the engines your buyers use, capture the verbatim responses, and record mention rate, citation rate, share of voice, sentiment, factual accuracy, and recommendation strength. Do not act on anything yet. The goal of this month is a defensible baseline and an accurate picture of where you actually stand.
Days 31 to 60, act on what the baseline reveals. Two findings usually dominate. If the engines are describing you inaccurately, that is the first priority, because a confident wrong answer actively costs you deals, and the fix usually lies in the trust and clarity layers of the previous papers. If competitors are dominating share of voice on unbranded questions, work backward from what the engines are citing for them and address the readiness gaps that keep you out of that set. Fix, do not chase. Improve the underlying content, evidence, and trust rather than trying to reverse-engineer an engine's current preferences, which the research suggests will not hold.
Days 61 to 90, establish the cadence and the reporting rhythm. Re-run the identical probe set, compare against the baseline, and report the change without claiming causation. Decide the ongoing interval, monthly is typical for competitive categories and quarterly is often sufficient otherwise. Formalize the report so it is comparable period to period. For agencies, this is where the engagement converts from a project into a monitoring retainer with a strategic review attached.
Checklist
Probe set design
- Fifteen to twenty-five prompts reflecting real buyer questions
- Discovery intent covered, unbranded category questions
- Comparative intent covered, versus and alternatives
- Recommendation intent covered, advice with constraints
- Branded factual intent covered, to detect misrepresentation
- Competitor set identified in advance
- Probe set versioned and held stable across periods
Execution
- Multiple engines probed, not just one
- Native versus live-web mode recorded where both exist
- Verbatim responses captured, not only scores
- More than one run per prompt where feasible, given non-determinism
Metrics recorded
- Mention rate
- Citation rate
- Position when mentioned
- Share of voice against named competitors
- Sentiment
- Factual accuracy, with specific errors logged
- Recommendation strength, tracked separately from mention
- Cross-engine variance
Interpretation discipline
- Single runs treated as anecdotes, not signals
- No causal claims from uncontrolled changes
- Trend compared only against the same probe set
- Factual errors escalated as the highest-priority finding
- A defined re-measurement cadence exists
See where your own site stands
Run a free AuditSpark AI & GEO Readiness audit — score, executive summary, and the fixes that matter, in minutes.
Run a free audit →Next in the series → The Agency Playbook for GEO Services