How to Measure AI Search Visibility

By Reputation.blog Editorial

The short answer

You measure AI search visibility the way you would measure anything produced by a system you cannot inspect: with a fixed set of questions, asked under controlled conditions, repeated often enough that variation in the answers becomes part of the measurement rather than a surprise inside it. Concretely: decide in advance which questions matter (a pre-registered question set), run each several times on each platform you care about, record the platform, the exact prompt, the date, the full response, and every citation URL, and score those records consistently — whether your business was named, whether it was cited, which sources were used instead, and whether what the answer said about you was accurate. Repeat the identical exercise at 30, 60, and 90 days.

Everything else in this guide is discipline around that core. AI answers are not rankings: there is no stable position to track, and — as the repeat-run research cited below shows — asking the same question twice in one day often returns different sources. A method that ignores that instability produces numbers that look precise and mean little. A method built around it produces something more modest and more useful: a defensible record of how AI systems describe and source your business over time.

This guide covers the full working method: what “visibility” consists of, eligibility checks to run before measuring anything, the step-by-step baseline procedure, honest citation share, why answers vary and the sampling discipline that forces, competitor benchmarking, the 30/60/90-day cycle, what platforms report natively (and what they do not), the limits of monitoring tools, and — stated plainly — which changes you can and cannot attribute to your own actions. For background on citations and how they differ across platforms, see our guide to AI citations.

What “AI visibility” actually consists of

“AI visibility” is used loosely in the industry to mean at least four different things, and a measurement program has to keep them separate because they behave differently and are recorded differently.

1. Mentions. Your business, brand, or product is named in the text of an AI-generated answer. A mention requires no link and no source: a model can name your business from patterns in its training data without retrieving anything. Mentions measure whether the system knows of you and includes you.

2. Citations. A platform displays a link or reference to a specific page alongside its answer, presented as a source for it. Citations are the observable footprint of retrieval: a page was found, read or indexed, and used. A citation may point to your own site or to a third-party page about you — different things, counted separately.

3. Description accuracy. What the answer says about your business — the services, locations, prices, dates, and claims it attributes to you — and whether those are correct. An answer can cite you and describe you wrongly. Visibility without accuracy is not an asset.

4. Downstream traffic and enquiries. Clicks and conversions that result — the layer businesses ultimately care about and the least directly observable: Google folds AI-feature clicks into standard Search Console reporting with no separate breakdown, and no AI platform gives ordinary businesses a per-answer referral report. Traffic is supporting evidence, not the primary metric.

A serious program tracks the first three directly and treats the fourth as context. Collapsing them into one score — as many vendor dashboards do — hides exactly the distinctions you need when deciding what to fix.

The platforms do not measure the same way, because they do not work the same way

Any method has to respect documented platform differences rather than averaging them away:

• Google AI Overviews and AI Mode are features of Google Search. Google documents that both may use a “query fan-out” technique — issuing multiple related searches across subtopics to build a response — and that a page is eligible to appear as a supporting link if it is indexed and eligible to be shown in Search with a snippet; there are no additional technical requirements. AI Overviews also do not appear for every query: Google states they show only when its systems judge them additive to classic Search, and “often don’t trigger.”
• ChatGPT answers some questions from model knowledge and others using search. OpenAI’s crawler documentation separates the machinery: OAI-SearchBot surfaces websites in ChatGPT’s search features (sites opted out “will not be shown in ChatGPT search answers”), GPTBot collects content that may be used for training, and ChatGPT-User fetches pages at a user’s request and “is not used to determine whether content may appear in Search.” A brand can therefore be mentioned from training-derived knowledge while having no presence in ChatGPT’s search layer at all.
• Perplexity is built around live retrieval: answers are assembled from web sources and presented with numbered citations. Perplexity documents PerplexityBot as the crawler that surfaces and links websites in its search results (and states it is not used to train foundation models), and Perplexity-User as a user-triggered fetcher that generally ignores robots.txt.
• Gemini and Microsoft Copilot are separate products with their own retrieval arrangements and citation behavior. Do not assume that what is documented for Google Search’s AI features describes Gemini’s standalone app, or that Bing’s index role in Copilot makes Copilot’s answers a readout of Bing rankings. Where a platform has not published its mechanism, this guide treats it as unpublished — because it is.

The practical consequence: measure each platform as its own dataset. A combined cross-platform “AI visibility score” is a reporting convenience, not a measurement.

Before you measure: eligibility checks

Measuring the visibility of a business that cannot be retrieved is measuring zero, repeatedly. Run these checks first and record the results as part of the baseline:

1. Indexation in the underlying indexes. For Google’s AI features, confirm key pages are indexed (Search Console’s Page indexing report and URL Inspection tool) — Google documents indexed, snippet-eligible status as the eligibility bar. For ChatGPT and Perplexity there is no equivalent webmaster console for most businesses; crawlability is the checkable layer.
2. Crawler access. Check robots.txt and any firewall/WAF rules against the documented agents: Googlebot (Google Search, including its AI features), OAI-SearchBot (ChatGPT search surfacing), PerplexityBot (Perplexity search results), and, for training-layer knowledge, GPTBot. OpenAI and Perplexity publish their crawler documentation and IP ranges, and both note robots.txt changes can take around 24 hours to be reflected. A blocked search crawler is a complete, mechanical explanation for absence from that platform’s citations.
3. Server logs. Confirm from your own logs whether those crawlers are actually visiting. Access granted and visits observed are different facts; record both.
4. Your native reporting baseline. If your site receives impressions in Google’s AI features, Search Console’s Generative AI performance report will show them. Note the starting values — the one platform-supplied time series you have.
5. Factual ground truth. Write down, from primary sources you control, the facts an accurate answer about your business should contain: correct name, locations, services, founding facts, product names. You cannot score description accuracy later without a reference version recorded now.

Only after these checks does a baseline mean anything: you will know whether absence in answers reflects ineligibility, or something more interesting.

The methodology, step by step

Step 1 — Build the question set, and freeze it

The question set is the instrument. Everything downstream depends on its quality.

• Write questions the way buyers ask them, in natural language: category questions (“best [service] in [town]”), comparison questions (“[you] vs [competitor]”), factual questions about your business (“does [business] offer [service]”), problem questions (“how do I [job your service does]”), and your name alone, for purely branded answers.
• Keep most of the set unbranded. If every question contains your name, you are measuring whether AI can talk about you when prompted with you — not whether you surface when a buyer describes a need.
• Size it for repetition, not breadth. Every question must be run multiple times per platform per round, so 20–40 well-chosen questions, run properly, beat 300 run once. Breadth you cannot repeat is anecdote at scale.
• Pre-register it. Write the set down, date it, and version it — the same discipline our methodology applies to client research. The set is fixed before results are seen, so later rounds cannot quietly swap in questions that flatter the trend. If you change the set later, that is a new baseline, and the record should say so.

Step 2 — Fix the conditions, and record them

For each run, record:

• Platform and surface (e.g., ChatGPT web app with search, Perplexity, Google AI Overview, Google AI Mode) — app, API, and search-results page are different surfaces and can return different sources.
• Date and time.
• Whether the session was logged in or logged out, and any personalization the account carries (memory, prior conversations, custom instructions). For a clean panel, use a fresh, logged-out or neutral session where the platform allows it, and say what you used.
• Location context, where the platform exposes or infers it. Local questions answered from the wrong geography measure the wrong market.
• The exact prompt text, verbatim. Paraphrasing between rounds invalidates the comparison.

Step 3 — Run repetitions: sampling discipline

Independent repeat-run research shows how unstable a single answer is. In Muck Rack’s What Is AI Reading? study (October 2026 edition), the same prompt run twice in one day saw ChatGPT repeat only 41% of its cited domains, Gemini 45%, Perplexity 94%. An academic audit of AI product recommendations found the same pattern from the other side: mean overlap between repeated responses was 26.0% of domains for ChatGPT, 29.8% for Gemini, 45.9% for AI Overviews — and repeated requests kept surfacing new sources, so one response captures only part of the source pool a platform draws on for that question.

The methodological consequences:

• Never report a single run. A prompt checked once is an anecdote with a timestamp.
• Run each question multiple times per round — three runs is a practical minimum; five is better for questions that drive decisions. Report presence as a rate across runs (“cited in 7 of 15 runs”), not as a yes/no.
• Expect platform differences in stability. A platform with high same-day repeat rates needs fewer repetitions to characterize than one with low rates; your baseline will show each platform’s behavior on your questions, so set repetition counts per platform.
• Spread runs within a round across a short window (the same few days) so a round describes a period, not an instant — but keep rounds on a fixed cadence (Step 8).

Step 4 — Record the full response, not just the score

For every run, save the complete record: platform, surface, prompt, date and time, session conditions, the full answer text, and every citation URL displayed, in order.

Full records matter because scoring decisions get revisited. When a definition changes or a hard question arrives, the raw record is the only thing that lets you recompute honestly. A spreadsheet of percentages with no underlying responses is not research data; it is a claim.

Step 5 — Code mentions, citations, and accuracy separately

Score each recorded answer on three independent axes:

• Mention: Is your business named in the answer text? Note where (first-named carries different weight in a recommendation list than last) and in what role — recommended, listed neutrally, compared unfavorably, or mentioned as something else entirely (a namesake, a former entity).
• Citation: Is any displayed source a page on your domain? Is any a third-party page about you? Record the exact URLs. A citation of a directory profile about you is evidence about the directory’s retrievability as much as yours — it shows where the platform prefers to source facts about businesses like yours.
• Accuracy: Against the ground-truth sheet from your eligibility checks, mark the answer’s statements about your business as accurate, inaccurate, outdated, or unverifiable. Record which claim was wrong; “the answer was inaccurate” is not actionable, “the answer lists a service discontinued in 2024, sourced from [page]” is.

Also record every competitor named and every source cited — Step 7’s raw material, captured free.

Step 6 — Compute the metrics, with the denominators showing

Four metrics cover most honest reporting. Each is simple; the honesty is in declaring exactly what was counted.

Presence rate (per platform). Of all runs in the round, in what share was your business mentioned at all? Presence rate = runs mentioning you ÷ total runs — the closest equivalent to “how often do we show up,” reported per platform because trigger and retrieval behaviors differ.

Citation rate (per platform). Of all runs, in what share was at least one of your own pages cited? Usually lower — often much lower — than presence rate, because answers frequently name businesses sourced from third-party pages.

Citation share. Your share of the citations themselves: citations pointing to your domain ÷ all citations displayed across the same runs. Report it alongside the shares of the third-party sources about you and of competitors’ domains.

Two warnings belong with this metric wherever it appears. First, there is no settled industry standard for citation share. Vendors define it differently: some divide your citations by all citations in the answer set; others divide answers citing you by answers that contain any citations; others restrict the denominator to tracked competitors’ domains. The same dataset yields materially different numbers under each. Second, a citation share figure is therefore only meaningful with its formula attached. When you see the metric quoted without one — in any sales deck, including a monitoring vendor’s — you are looking at a number you cannot reproduce, compare, or audit. In your own reporting, print the formula next to the number, every time.

Mention share (share of voice). Your mentions as a share of all tracked brands’ mentions in the same runs: mentions of you ÷ (mentions of you + mentions of each tracked competitor). Like citation share, this is a family of definitions, not one standard — some count each answer once per brand, others count every mention occurrence. Pick one, document it, never change it mid-series.

A note on precision: with runs in the low hundreds, these percentages carry wide uncertainty — a move from 22% to 27% across rounds may be pure noise. Report counts alongside percentages (“34 of 150 runs”), and treat small movements as noise until they persist across rounds.

Step 7 — Benchmark competitors on the same instrument

Competitor benchmarking is valid only because the instrument is shared: competitors are scored from the same runs of the same questions under the same conditions — what makes the comparison fair in a way that separate per-brand checks never are.

Choose a small set of real competitors — the businesses a buyer would genuinely weigh against you — and fix the list in the pre-registration. From each round, report for each competitor the same presence rate, citation rate, and citation share you compute for yourself. Then read the source data, not just the scores:

• Who is cited when you are not? The pages platforms use to answer your category’s questions — directories, review platforms, news coverage, a competitor’s own guide — map where retrievable authority in your category currently sits. This is often the most commercially useful output of the exercise: it shows which sources to earn, correct, or build — the subject of our AI visibility pillar.
• Whose facts are used? If answers describe the category using a competitor’s framing, prices, or terminology, that is a positioning finding, not just a visibility finding.
• Where are you named but not cited? That gap usually means platforms know you from training data or third-party text but do not treat your own pages as the source — a specific, addressable condition.

Step 8 — Re-measure on a 30/60/90-day cycle, and keep an interventions log

Visibility measurement is a time series or it is nothing. Fix the cadence in advance: same question set, same conditions, same scoring, re-run at 30, 60, and 90 days from baseline, and quarterly thereafter if the program continues.

Alongside the rounds, keep a dated interventions log: every change you make that could plausibly affect results (pages published or revised, facts corrected on third-party profiles, entity information made consistent, crawler access changed), and every external event you become aware of (platform model updates, changes to a platform’s search features, a competitor’s campaign, press coverage). The log separates a time series you can reason about from a line on a chart that invites a story; how to reason with it — and where reasoning must stop — is the attribution section below.

A worked example (illustrative data)

The following example uses invented, illustrative numbers to show the computation. It is not a real business, a real client, or real results.

Suppose a regional services business — call it Example Co — pre-registers 20 questions (5 branded, 15 unbranded) and runs each 3 times on two platforms in its baseline week: 120 recorded answers. Baseline coding produces, per platform (60 runs each):

Measure	Platform A	Platform B
Runs mentioning Example Co	18 (30%)	27 (45%)
Runs citing an Example Co page	6 (10%)	15 (25%)
Total citations displayed across all runs	214	236
Citations to Example Co’s domain	9	19
Citation share (own domain ÷ all citations)	9/214 = 4.2%	19/236 = 8.1%
Citations to third-party pages about Example Co	17	22
Runs mentioning top competitor	33 (55%)	30 (50%)

Read honestly, this baseline says: Platform B names and cites Example Co substantially more than Platform A; on both platforms, third parties are cited about Example Co more than Example Co’s own site is; and the leading competitor is named in about half of all runs. Had the business instead defined citation share as “answers citing us ÷ answers containing any citation,” Platform A’s figure would have come out far higher — precisely why the formula travels with the number.

At the 30-day re-measurement, suppose Platform A presence rises from 30% to 37% (22 of 60 runs), with no intervention logged and no known platform change. The correct reading is not “visibility improved 7 points” but “the rate moved by an amount small relative to this panel’s run-to-run variation; note it and wait for the 60-day round.” The method obliges you to say exactly that.

What the platforms report natively

Your own panel is the primary instrument, but two platform-supplied sources belong in every program.

Google Search Console — Generative AI performance report. Google documents this report as showing impressions — how many times links to your site were shown in a generative AI feature on Google Search — covering AI Overviews and AI Mode together, by page, country, device, and date; Google states it was rolled out to all websites worldwide as of August 31, 2026. What it omits matters as much: no clicks, no queries, no separation of AI Overviews from AI Mode. It is a genuine platform-measured time series of exposure — but not of which questions produced it or what anyone did next.

Google Search Console — Performance report. Google documents that sites appearing in AI Overviews and AI Mode are included in overall Search traffic, reported within the “Web” search type. Clicks from AI features sit inside your standard clicks, impressions, and position data, inseparably mixed with classic results. Anyone claiming to show you “your AI Overview clicks” from Search Console alone is showing you an estimate wearing a uniform.

Everything else is your own instrumentation. Neither OpenAI nor Perplexity offers ordinary businesses a webmaster report of how often their pages were cited in answers. Your remaining native signals are indirect: analytics referrals from AI platforms (clicks from ChatGPT, Perplexity, and similar surfaces appear as referrals or campaign-tagged traffic, though much AI-influenced traffic arrives with no usable attribution), and server logs showing crawler visits — evidence of eligibility and indexing interest, not of answers given.

Monitoring tools and their limits

A market of AI visibility monitoring tools (among them Profound, Peec AI, and the AI toolkits of established SEO platforms) now automates the panel method: scheduled prompt runs, captured answers and citations, charted presence and share metrics. Used well, a tool buys scale and consistency. Used uncritically, it buys false confidence. The limits to weigh:

• The surface problem. Tools collect answers through APIs, automated browser sessions, or partnerships — not through the exact consumer interface your customers use. Independent research has found that a provider’s API and consumer interface return substantially different source sets (in the arXiv audit cited earlier, mean domain overlap between ChatGPT’s interface and its API view was 12.0%; for Gemini, 14.8%). A tool measures a surface. Ask which one, and spot-check it against manual runs.
• The definition problem. Headline metrics — citation share, share of voice, visibility score — are defined differently by different vendors. Before adopting a tool’s dashboard as your reporting, reproduce one of its figures from its own raw answer data.
• The sampling problem. A tool’s cadence and run counts set the noise floor of every chart it draws. If it runs each prompt once per check, its trend lines are built from the same single-run instability the research warns against, at industrial scale. Ask how many repetitions sit behind each data point.
• The personalization and location problem. Automated collection runs from the tool’s accounts, locations, and contexts, which may not resemble your buyers’. And tools only track the questions you give them: they cannot observe the questions you did not think to ask, or score description accuracy against your ground truth.

The sound arrangement: let a tool carry the volume if budget allows, but keep a smaller manual panel, run by hand each round, as the calibration standard. When tool and panel disagree, believe the records you can inspect.

Attribution: what changed, and what you may claim about it

What movement in your series can establish: that AI systems’ answers about your category changed between rounds, in specified directions, under specified conditions. That is a real and useful finding.

What it cannot establish by itself: that your actions caused the change. Between any two rounds, the following also move: the platforms’ models and retrieval systems (updated on their schedule, usually without notice), their indexes, your competitors’ content and coverage, press and directory content about everyone in the category, and market demand itself. A rise after your intervention is correlation in time — consistent with your intervention having worked, and equally consistent with several things you did not do.

Discipline that narrows — without eliminating — the gap:

• Log interventions and platform events with dates (Step 8), so explanations can be checked against timing. A change that precedes your intervention did not come from it.
• Watch for mechanism-shaped evidence. If answers begin citing the specific page you published, or quoting the specific facts you corrected, that is stronger than a bare rate increase, because the content of the change matches the intervention. Still not proof: platforms re-crawl and re-index on their own.
• Use a comparison set. Keep a few panel questions your work did not target. If targeted and untargeted questions move together, a platform-wide change is the likelier explanation; if only targeted ones move, your case strengthens — to “consistent with,” not “proven.”
• State results in observational language. “Citations of our profile pages rose from 4% to 9% of citations on our panel over 90 days, beginning the round after the pages were published and the directory records corrected” is an honest sentence. “Our work increased our AI visibility by 125%” is a claim the underlying system does not let anyone verify.

Reputation.blog’s own commercial work is sold under the same constraint, stated here because this guide sits in our Knowledge Center: payment for AI visibility work covers research, optimization, content development, and publication. It does not purchase favorable editorial conclusions, and it does not guarantee citations, rankings, recommendations, or visibility on any platform — because no honest provider can. Any measurement report that reads like a guarantee has stopped being a measurement report. Our methodology page sets out how we keep research and commercial work separated.

Business implications

Measured properly, AI visibility data supports a short list of sound decisions:

• Where you are actually absent. Never being named on a category question set is a different problem from being named but inaccurately described — and both differ from being described well but sourced from a third party you’d rather correct. Each diagnosis points at different work: documentation, correction, or source development.
• Which sources run your category. The baseline’s citation map shows where buyers’ AI answers are sourced from today — for most businesses, the first time that map has existed.
• Whether change is real. The 30/60/90 series protects budget in both directions: it stops you crediting a vendor (or yourself) for noise, and stops you abandoning sound work because one spot-check looked bad.

It does not support ranking-style reporting to leadership (“we are #2 in ChatGPT”), week-by-week performance management of a system with this much variance, or any claim that a citation movement was caused by a specific intervention. The honest version of this discipline — audits against a recorded baseline, development work documented as it happens, monitoring on a fixed cadence — is what our AI Visibility Audit and AI Visibility Monitoring services are built on; the standards in this guide are the standards those services report against.

Common misconceptions

“I asked ChatGPT once and we weren’t in the answer, so we have no AI visibility.” One answer is one draw from a variable system; the same prompt re-run the same day can return a substantially different source set. Absence (and presence) only means something as a rate across runs.

“A mention and a citation are the same thing.” A mention is your name in generated text; a citation is a displayed source link. Answers routinely contain one without the other, produced by different mechanisms — training-derived knowledge versus live retrieval. Counting them together makes diagnosis impossible.

“Citation share is a standard metric, so I can compare our number to a benchmark.” There is no agreed standard definition; vendors compute it differently and no platform publishes an official version. Benchmark comparisons of citation share compare formulas, not performance.

“Search Console can show me my AI Overview traffic.” Google documents AI-feature traffic inside the Performance report’s Web search type with no separate breakdown, and its Generative AI performance report shows impressions only — no clicks, no queries. A clean “AI traffic” line in a report came from somewhere else, usually an estimate.

“The monitoring tool sees what our customers see.” Tools observe through APIs or automated sessions, from their own locations and contexts. Documented interface-versus-API divergence means a tool’s panel is a model of customer experience, not a recording of it. Calibrate against manual checks.

“Our citations went up after the work, so the work caused it.” Sequence is not causation in a system whose models, indexes, and sources change constantly. At best, timing plus mechanism-shaped evidence plus a comparison set makes your intervention the plausible explanation — and reports should say exactly that.

“If we’re not being cited, we’re invisible.” Platforms name businesses from training-derived knowledge without citing anything. You can be recommended and uncited — or cited and misdescribed. Hence the separate scoring.

Frequently asked questions

How many questions do I need in my set?
Fewer than you think, run more times than you expect. Twenty to forty well-chosen questions, each run at least three times per platform per round, teach you more than hundreds run once.

Can I do this without a monitoring tool?
Yes. The manual panel in this guide — a spreadsheet, a fixed question set, disciplined recording — is the complete method. Tools add scale and convenience, not validity, and the manual panel remains the calibration standard if you later adopt one.

What should a good baseline report contain?
The question set and its version date; the conditions of each run; presence rate, citation rate, citation share (with formulas), and accuracy findings per platform; the competitor table from the same runs; the eligibility results; and the full recorded responses, or a link to them. Missing any of those, the baseline cannot be re-run or audited — and cannot anchor a 90-day comparison.

Our numbers went down. Did we do something wrong?
Not necessarily. Check, in order: eligibility (crawler access, indexation), known platform changes, competitor or press events in the window, and whether the movement exceeds your panel’s normal run-to-run variation. A decline inside the noise band is a reason to watch the next round, not to reverse course.

References

Where this guide reports figures, they come from the named study, not from platform claims.

• Google Search Central — AI Features and Your Website (query fan-out; eligibility; AI-feature traffic reported in the Performance report within the “Web” search type): https://developers.google.com/search/docs/appearance/ai-features
• Google Search Console Help — Generative AI performance report (Search) (impressions for AI Overviews and AI Mode; worldwide rollout as of August 31, 2026): https://support.google.com/webmasters/answer/16984139
• OpenAI — Overview of OpenAI Crawlers (OAI-SearchBot for search surfacing; GPTBot for training; ChatGPT-User for user-initiated fetches): https://developers.openai.com/api/docs/bots
• Perplexity — Perplexity Crawlers (PerplexityBot surfaces and links sites in search results; Perplexity-User for user actions): https://docs.perplexity.ai/guides/bots
• Muck Rack / Generative Pulse — What Is AI Reading?, October 2026 edition (same-prompt repeat rates for cited domains: ChatGPT 41%, Gemini 45%, Perplexity 94%): https://muckrack.com/blog/what-is-ai-reading-october-2026
• Academic audit of AI-generated product recommendations (repeated-request domain overlap: ChatGPT 26.0%, Gemini 29.8%, AI Overviews 45.9%; interface-versus-API divergence), arXiv: https://arxiv.org/html/2609.18729
• Writesonic — Brand Explorer documentation (one vendor’s published definitions for visibility, share of voice, and citation share — cited to show definitions are vendor-specific, not standard): http://docs.writesonic.com/docs/brand-explorer

---

Published: October 31, 2026 / Last reviewed: October 31, 2026
Attribution: Reputation.blog Editorial