How to measure AI visibility in SEO - Ighenatt Blog

AI visibility measurement has a neat trap: it looks easy until you repeat the same query twice. On Tuesday ChatGPT cites your guide. On Wednesday it disappears. Perplexity links to you in the first answer, then swaps two sources while keeping the same conclusion. Google shows an AI Overview on mobile, but not on desktop. If your dashboard turns that variability into a single number, the number looks serious and the decision goes sideways.

The contrarian point is this: you do not need one more metric, you need a sampling methodology. The resource on GEO metrics already covers the KPIs. The guide to GEO monitoring tools compares platforms. This article goes into the uncomfortable part: how to design the experiment so a brand can say, cautiously, “we have AI visibility” without mistaking a screenshot for evidence.

First define what you are measuring

AI visibility does not mean the same thing in ChatGPT Search, Gemini, AI Overviews and Perplexity. Google explains in its AI features documentation that AI Overviews and AI Mode may use query fan-out, issuing related searches to build an answer with supporting links. OpenAI documents that ChatGPT Search may show inline citations and a Sources panel. Perplexity describes its product as searching the web in real time and supporting each answer with numbered citations.

That interface difference changes measurement. In Google you can observe impressions and clicks with Search Console, but you still need manual sampling or an external tool to know whether a URL was used as a supporting link in a specific answer. In ChatGPT Search you can record citations when they appear, but there is no OpenAI Search Console for publishers. In Perplexity the citation is more visible, though it also varies by model, time, language and wording.

The minimum unit of analysis should not be “keyword”, but “prompt observed in context”. A prompt includes the exact wording, language, country, device when relevant, engine, date, time, search mode, account used and whether history was clean. Yes, it sounds bureaucratic. It also sounds like how you avoid spending weeks optimizing for a false positive.

According to Ronald Sielinski, author of “Quantifying Uncertainty in AI Visibility”, citation metrics should be treated as sample estimators of an underlying response distribution, not fixed values. That is the mental shift. In classic SEO we accept average position; in AI we need to accept ranges, intervals and noise.

Design a prompt corpus, not a keyword list

A useful corpus combines real demand with conversational variation. The practical approach is to start with 30 to 80 prompts, grouped by intent. For a B2B SaaS, for example, “best reporting software” is not enough. Include “what tool should an agency use for SEO reporting”, “Looker Studio alternatives for clients”, “how to measure visibility in ChatGPT” and “software to monitor AI mentions in Spanish”.

The baseline sampling structure we use in advanced audits looks like this:

The sample should cover head, mid-tail and long-tail prompts. Head prompts provide demand, but long-tail prompts reveal which sources the engine uses when it needs precision. To avoid writer bias, pull candidates from Search Console, sales questions, internal site search, People Also Ask, form logs and real customer support queries. Then remove semantic duplicates and save the corpus with stable IDs: P001, P002, P003.

Do not change the corpus every week. Keep 80% as a fixed panel for trend analysis and 20% as an experimental panel for new questions. Put differently: the corpus is a lab bench. You can move samples around, but do not replace the bench every time you want comparable results.

Repeat runs and calculate uncertainty

The most common bad habit in AI visibility is measuring once and presenting a percentage. “We appear in 12 of 40 prompts” sounds clear. But if each prompt was run only once, you do not know whether you have a stable signal or a lucky draw.

The 2026 papers “Quantifying Uncertainty in AI Visibility” and “Don’t Measure Once” reach the same operational conclusion: generative answers vary across runs, prompts and time. The first study analyzed Perplexity Search, OpenAI SearchGPT and Google Gemini with daily and high-frequency sampling; the second argues for characterizing visibility as a distribution rather than a point observation.

For SEO teams, a reasonable design is:

1. Run each prompt 5 times per engine in the initial audit.
2. Spread runs across at least 3 days when budget allows.
3. Repeat critical prompts weekly and the full corpus monthly.
4. Calculate observed proportion, confidence range and change versus the previous period.

If P014 is run 10 times in Perplexity and your domain is cited 6 times, report a 60% observed citation rate, not “ranking won”. If next month it appears 7 out of 10 times, do not sell it as an automatic victory. With small samples, the difference may sit inside the noise. Bootstrap intervals or Wilson intervals help avoid overexcited conclusions.

The report should show three states: stable presence, intermittent presence and absence. Intermittent presence is the most actionable. It means the engine already considers your domain a candidate, but does not choose it consistently. That is where improving citable passages, external authority, data freshness and internal links toward the candidate page makes sense.

Separate citations, mentions, impressions, clicks, referrals and conversions

Putting everything into an “AI visibility score” is convenient. It also erases decisions. A citation is not a mention; an impression is not a click; a referral is not a conversion. Each layer answers a different question.

Citations: record whether the engine links to your domain or URL as a source. In Perplexity this is usually a numbered citation. In ChatGPT Search it may appear inline or in the Sources panel. In Google AI Overviews it may appear as a supporting link. Citation measures selection as a source.

Mentions: record whether the answer names the brand, product, author or entity without linking. Ahrefs separates mentions from citations in its AI visibility tool, and that distinction is healthy: a brand may be recommended without receiving a link.

Impressions: in Google, Search Console defines impressions as times someone saw a link to your site on Google. Do not extrapolate that definition to ChatGPT or Perplexity if the platform does not provide impressions. For engines without publisher data, use “observed exposures in sample”, not real impressions.

Clicks: Search Console measures clicks from Google Search. Google also recommends analyzing conversions and time on site in tools such as Analytics to understand traffic quality from results with AI Overviews. Outside Google, clicks only appear if they arrive with an identifiable referrer or UTM.

Referrals: in GA4, review Session source / medium and Page referrer for domains such as chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com or app variants. The Google Analytics 4 SEO guide explains how to cross acquisition, engagement and conversion without mixing scopes.

Conversions: define business events for AI traffic: form submission, booking, download, demo, newsletter, call or visit to a service page. A citation that does not bring a click may still influence a later branded search. That is why you should also watch branded search growth and assisted conversions, without attributing everything to AI by magic.

Cross GSC and GA4 without inventing attribution

Google Search Console is the baseline for Google, not for the whole AI ecosystem. Its documentation defines four main metrics in the Performance report: clicks, impressions, CTR and average position. It also warns that results can vary by time, place, device and search history. That nuance matters when you compare a manual screenshot with aggregated data.

The correct workflow starts in GSC: identify pages with many informational impressions, low CTR and potential presence in AI Overviews. Then run the related prompt corpus and mark which URLs appear as supporting links. Finally, take those URLs into GA4 and compare engagement, conversions and session source.

Avoid three mistakes. First: attributing a CTR drop to AI Overviews without checking ranking changes, seasonality, snippets, competitors or intent. Second: assuming every chatgpt.com / referral visit comes from a measurable citation; it may come from a link pasted into a private conversation. Third: declaring “AI traffic” by looking only at referrals, because many users see a brand in an AI answer and then search the brand on Google or enter directly.

A good monthly table has four blocks: observed visibility by prompt, GSC performance for candidate URLs, GA4 referrals from AI platforms and subsequent conversions. When those four blocks move in the same direction, you have a story. When they do not, you have hypotheses.

Use third-party tools with statistical discipline

AI visibility tools save work, especially when you monitor hundreds of prompts or several competitors. Semrush documents metrics such as AI Visibility Score, mentions, citations, cited pages, sources and missing prompts, with coverage across Google AI Overviews, AI Mode, Gemini and ChatGPT in its Visibility Overview. Ahrefs says it queries ChatGPT, Gemini, Perplexity, Copilot and AI Overviews with prompts derived from search behavior and separates mentions from citations.

That does not remove the methodology question. Before accepting any score, ask or review:

• Which prompts the tool uses and whether they are real, synthetic or mixed.
• How many runs it performs per prompt and how often.
• Which country, language, device and engine it covers.
• How it identifies citations versus mentions.
• Whether it stores full answer, sources, timestamp and cited URL.
• Whether it exports raw data so you can recalculate metrics.

Profound, Otterly, Peec AI, Semrush, Ahrefs and similar tools can be part of the stack. But the tool does not decide meaning. If a provider does not show uncertainty, document your own qualitative margin: stable, volatile or exploratory. (Yes, it is less spectacular than a 0 to 100 score. It is also more honest.)

Limits you should write into the report

AI visibility measurement has structural limits. First, website owners do not have complete impression logs from ChatGPT, Gemini or Perplexity. Second, engines change models, interfaces and citation policies without warning your dashboard. Third, personalization, location and history can alter responses. Fourth, citations do not always perfectly support the associated sentence: the Liu, Zhang and Liang study on verifiability found support and accuracy issues in generative search engines.

There are business limits too. More mentions may not bring traffic. Fewer referrals can coexist with more branded searches. A Perplexity citation may be valuable in technical B2B and marginal for a low-ticket ecommerce store. The methodology must end in decisions: which content to update, which entities to strengthen, which pages to turn into citable sources and which topics to abandon because the signal is weak.

My preferred audit ending is a simple matrix: maintain, strengthen, investigate, discard. Maintain prompts with stable presence and conversions. Strengthen prompts with intermittent presence and strong intent. Investigate prompts with many mentions but zero clicks. Discard prompts with no presence, no demand and no clear commercial relationship.

The next practical step: create a corpus of 40 prompts, run it 5 times in three engines, tag citations and mentions separately, cross those URLs with GSC and review referrals/conversions in GA4. In two weeks you will have an imperfect but useful baseline. Much better than a shiny screenshot and no decision.

FAQ about AI visibility measurement

How many times should each prompt be repeated to measure AI visibility?

For an initial audit, use at least 5 runs per prompt and engine. For critical prompts, increase to 10-15 runs spread across several days. There is no magic number: the goal is to detect whether presence is stable, intermittent or accidental.

Does Google Search Console show AI Overview citations?

Search Console measures clicks, impressions, CTR and position for links in Google Search. Google says AI Overviews and AI Mode are part of Search, but GSC does not replace a citation and mention measurement system for ChatGPT, Perplexity or Gemini.

What is the difference between an AI citation and a mention?

A citation links to or attributes a URL as a source. A mention names a brand or entity without necessarily linking. For SEO, measure both separately because they have different effects: potential traffic, perceived authority and brand recall.

Which tools can measure AI visibility?

Semrush, Ahrefs, Profound, Otterly, Peec AI and other platforms help automate prompts, citations, mentions and competitive benchmarking. Use them alongside your own corpus and GSC/GA4 data so you are not dependent on an opaque score.