How to Measure AI Visibility
AI visibility means tracking whether assistants name and link your brand across a fixed set of buyer questions run weekly, not a single good screenshot.
AI visibility means tracking, on a fixed schedule, whether assistants like ChatGPT, Claude and Perplexity name your brand and link your specific pages when a real buyer question is asked. It is measured as a rate across many questions and many weeks, never as a single good answer.
That measurement discipline is the part of generative engine optimization most teams skip, which is exactly why it is worth doing properly rather than treating a lucky screenshot as proof of anything.
في هذه الصفحة
What does "AI visibility" actually mean?
It means the rate at which assistants surface your brand when someone asks a question your buyers would plausibly ask, tracked consistently over time. That is a narrower and more useful definition than "how often does AI mention us," because it forces a fixed, repeatable question set rather than an unstructured habit of occasionally typing something into ChatGPT and reacting to whatever comes back.
The tracking study behind much of the current data on this topic logged over thirteen thousand individual citations across four major assistants to build a picture of how each one behaves [1]. That is the scale a real measurement exercise operates at — not one question, but many, run repeatedly, aggregated into a rate.
What should you actually track: mentions, citations, or both?
Both, and treat them as different outcomes rather than folding them into one score. A mention is the assistant naming your brand somewhere in its answer text. A citation is a link back to a specific URL of yours. An assistant can name you with no link, or link a third-party page that discusses you without naming your own site at all.
| Signal | What it tells you |
|---|---|
| Named, with a link to your site | The strongest outcome: your own page is the source |
| Named, with no link at all | You are known, but no page of yours was quotable enough |
| Not named, but a page about you is cited | A third party is doing the work your site should be doing |
| Not named, no related citation at all | A visibility gap worth investigating from scratch |
The gap between "named" and "linked" is itself a useful diagnostic, not just a vanity distinction. If you are consistently named but rarely linked, the fix is usually on your own pages: sharper, more citable, more specific facts, so an assistant has something worth pointing to. If you are rarely named at all, the fix is more often off-site entirely: third-party profiles, comparisons and mentions that simply do not exist yet, no matter how good your own pages already are.
Record both numbers every week, side by side, rather than collapsing them into one score. A single combined metric hides which of the two very different problems you actually have, and the fix for each one is different enough that guessing wrong wastes real time.
How do you build a fixed prompt set that is actually useful?
Start from real buyer language, not keyword-research phrasing. Twenty to forty questions is a workable range: enough to reveal a genuine pattern across a category, few enough to run by hand every week without a paid tool. Write them the way a person actually types into a chat box, including the vague, under-specified ones, because those are common and an assistant handles them differently than it handles a precise query.
Write questions in plain buyer language
"best tool for X" and "is X worth it" behave differently from a clean keyword phrase, and both are worth including.
Include your brand name in some, omit it in others
Both patterns happen in real usage, and an assistant can behave differently depending on whether the question already names you.
Cover the whole buying journey, not just "best X"
Pricing, comparison, and "is it worth it" questions surface different sources than a pure discovery query does.
Keep the list fixed once you start
Rewriting the questions after a bad week resets the only baseline the exercise was built to produce.
That kind of large-scale tracking study is where most of the current published benchmarks for this cluster come from [1], but your own thirty-question weekly spreadsheet is what actually tells you where your brand stands, not a benchmark built from someone else's dataset.
What tools actually do this tracking for you?
A growing category of paid tools now runs prompt sets like this automatically across several assistants and reports a visibility or share-of-voice score, and industry coverage of the category describes the general approach as running fixed queries on a schedule and aggregating the results, which is the same method described above at smaller scale [2]. The tools mostly add automation and cross-competitor comparison, not a fundamentally different method.
Start manually before paying for one. A spreadsheet with your question list, run weekly by a real person copying answers into rows, teaches you what the tools are actually measuring and where their categorization can go wrong, which makes you a much better buyer of one later if you decide the manual version does not scale.
Who should actually own this inside a company?
Usually whoever already owns organic search, with one adjustment: this work needs a fixed weekly slot on a calendar, not an ad hoc check whenever someone remembers. Treating it as a standing recurring task rather than a one-off audit is the single biggest predictor of whether a measurement programme like this survives past its first month, because the value only shows up once you have several weeks of data to compare against each other.
Marketing teams sometimes assign this to whoever is newest or has the most spare time, which tends to produce inconsistent question sets and gaps in the weekly record. It works better assigned to one specific person with the authority to say no to changing the prompt list on a whim, because protecting that list is most of the discipline this whole exercise depends on.
The programmes that work are the boring ones: same thirty questions, same day every week, same spreadsheet, for months. The moment someone starts swapping questions to make the numbers look better, the data stops meaning anything.
What should the report actually say each month?
More than a single trend line. A citation rate going up or down tells you the direction, not the reason, and the reason is what a stakeholder actually needs to hear. Pair the number with one sentence naming what changed that month: a new profile went live, a crawler block got fixed, a competitor started running a campaign that pushed their own mentions up.
Six months into a programme like this, that monthly sentence is often the only real record of what was tried and what happened as a result. A chart with no accompanying explanation ages badly, because nobody remembers, a year later, whether a dip in March was a real problem or just a bad week in an inherently noisy measurement. Write the sentence at the time, not in retrospect.
How often should you run the check, and what counts as good?
Weekly, at minimum. AI Overviews vs AI Mode already covers why the same query can trigger a different feature, or a different set of sources, on different days; that variance holds across every assistant in this cluster, not just Google's, which is why a single run of your prompt set is closer to noise than to a result.
There is no universal benchmark for what counts as a good rate, because the category is new and behavior varies enormously by industry and by how competitive a specific question is. A niche B2B category with three real competitors will naturally show a far higher citation rate than a crowded consumer category with hundreds of similar options, and comparing the two numbers directly tells you nothing useful about either one.
Treat your own first month as the baseline, and judge every month after against that baseline rather than against a number a vendor markets or a competitor claims in a press release. If you are also investigating why your brand is not in ChatGPT specifically, this same weekly tracking is the tool that tells you whether a fix actually worked, rather than guessing from memory or reacting to one conversation that happened to go well.
أسئلة يطرحها الناس
What is the difference between a mention and a citation?
A mention is the assistant naming your brand in its answer text. A citation is a link back to your specific URL. Tracking both matters, because the gap between them is itself useful information.
How many questions should I track?
Twenty to forty is a workable range for most businesses: enough to see a real pattern, few enough to run by hand every week without needing a paid tool immediately.
How often should I run the check?
Weekly. The same prompt run on different days can return different sources, so a single run is noise and only a trend across several weeks is signal.
Do I need a paid tool to do this?
Not to start. A spreadsheet and thirty minutes a week is a real measurement system. A paid tool becomes worth it once you are tracking across many assistants and many competitors at once.
What counts as a good citation rate?
There is no universal benchmark yet, because the category is too new and too fragmented by industry. Track your own trend against your own baseline rather than a number a vendor quotes.
المصادر
- I tracked 13,184 AI citations across 4 LLMs: here's what I learned — Orbit Media Studios (checked 23 September 2026)
- How to measure and improve AI visibility — Semrush (checked 23 September 2026)
- Generative engine optimization: a complete guide — Semrush (checked 23 September 2026)
Get your AI visibility baseline measured this week
احجز استشارة مجانية مدتها 30 دقيقة مع إبراهيم وجيريمي. ننظر في موقعك أثناء الحديث، وتخرج بخطة مرتّبة حسب الأولوية.
من كتب هذا
Jeremy ZavaCo-founder
Leads growth and client partnerships.
يعمل على: CRO, SEO
راجعه قبل النشر Ibrahem Helal، Co-founder.
المزيد عن هذا
- Generative engine optimization: the playbookHow answer engines pick what to cite, what a quotable page looks like, which crawlers to allow, how to measure it, and a 90-day plan built on evidence.
- AI Overviews vs AI Mode: What's the DifferenceAI Overviews sits inside normal Google results; AI Mode is a separate conversational search. Here is how they differ, and which to optimize for first.
- llms.txt Explained: Does It Actually Workllms.txt is a proposed file that lists a site's key pages for an AI model to read. No major assistant is confirmed to use it yet. Here is what it actually does.
- Why Your Brand Is Not in ChatGPT AnswersUsually a crawler block, a page nothing else links to, or simply no third party discussing you yet. Here is how to find which one it actually is.