AI visibility is whether your brand appears, is cited, or is recommended in AI answers to buyer questions. To measure it, repeat a fixed set of prompts on each priority engine and record brand mentions, linked sources, competitors, and answer accuracy. At #10, we separate those observations from referral traffic and inquiries, so a citation is not mistaken for a sale.
The useful question is not whether an assistant can repeat your company name when asked. It is whether a buyer asking about your category encounters your brand, and whether the answer gives them an accurate reason to consider it.
Our generative engine optimization guide explains the broader strategy. This guide focuses on the measurement: what counts, what to record, and how to compare checks without manufacturing progress.
Define what counts before collecting answers
A mention, a citation, and a recommendation are different observations. Combining them into one visibility score makes the report harder to act on.
| Observation | Counting rule we recommend | What it does not prove |
|---|---|---|
| Brand mention | The answer names your brand, allowing verified spelling variants | That it used your website as a source |
| Owned-site citation | A displayed source links to a page on your verified domain | That the answer recommends your product or service |
| Third-party brand reference | A linked external source discusses your brand; record the source and verify the passage | That your owned site earned the citation |
| Recommendation | The answer presents your brand as an option for the stated buyer need | That the buyer clicked or purchased |
| Accurate description | The answer gets the relevant product, service, audience, or limitation right | That the answer appeared for anyone beyond the observed test |
Open the cited links. Check the destination and the passage supporting the answer rather than accepting a familiar domain or a generated source label. If an assistant names your brand but supplies no verifiable link, record a mention, not an owned-site citation.
For a premium DTC brand, a product recommendation and a citation to its care guide are useful in different ways. For a B2B supplier, a technical article can be cited without the supplier making the shortlist. Keep both observations, but do not call them the same result.
Build a prompt panel around decisions buyers make
Start with the questions that could lead to a product choice, supplier shortlist, or inquiry. Our visibility framework groups prompts by use case, comparisons, category education, and technical buying advice. Choose the groups relevant to your business rather than testing every possible question.
Keep branded and unbranded prompts separate. Asking an engine about your company tests recognition and accuracy. Asking it to recommend a supplier without naming you tests whether you enter the consideration set. A report dominated by branded questions can look healthy while missing that second problem.
For each prompt, record its buyer intent and the page you would want cited. A DTC comparison question might map to a buying guide. A B2B application question might map to a technical category page with selection criteria. If there is no suitable page, that is a content gap to investigate, not evidence that publishing any new article will fix visibility.
Freeze the wording for the baseline. Add new questions as a separately labeled group instead of quietly replacing prompts where the brand performs poorly. Human judgment sets the direction here. Automating a panel of irrelevant questions only produces an irrelevant report faster.
Keep each engine and search surface identifiable
A Google AI Overview check is not a ChatGPT search check. Treat Google AI Overviews, Google AI Mode, ChatGPT search, and Perplexity as separate surfaces when they are in scope. Add other assistants when there is a buyer reason to monitor them.
For a practical evidence log, we recommend one row per prompt, surface, and test run, with these fields:
- Exact prompt and panel version, including whether the question names the brand.
- Engine, search or browsing mode, date, language, location, and any account or session conditions you can establish.
- Whether an answer or AI Overview appeared, plus any collection error.
- Brand mentioned, owned-site URL cited, third-party references, and recommendation status.
- Competitors and cited URLs, with the relevant answer passage or a saved capture.
- Accuracy problems, missing information, and a candidate page or action to review.
Use fresh conversations for comparable standalone checks. If you deliberately test a follow-up conversation, label it separately. Otherwise, a recommendation prompted by earlier discussion can be mistaken for an unaided discovery.
Repeat important checks under the same documented conditions. Preserve differing answers rather than keeping only the one where your brand appears. This is a sampled view of a surface, not a census of every answer buyers receive.
Calculate citation rates with an explicit denominator
For a starter scorecard, we recommend the following definitions. They are reporting conventions, not a universal industry scoring standard.
Mention rate: valid answer observations naming your brand divided by all valid answer observations in that panel and surface.
Owned-site citation rate: valid answer observations linking to your domain divided by all valid answer observations in the same panel and surface.
Count each answer once for these rates, even if it links to several of your pages. Keep the individual URLs in the evidence log so you can see which pages earn citations.
Google AI Overviews need an additional distinction. Track how often an Overview appears across successfully checked searches, then report owned-site citation rate among the observed Overviews. A successful search with no Overview is a surface-absence observation, not a failed collection and not an Overview that omitted you. A timeout or blocked check belongs in a separate error count, outside the valid-result denominator.
Always show the counts alongside the rate. An improvement based on a small panel deserves a different level of confidence from a pattern that repeats across many observations. If there are no valid answers or no observed Overviews, the corresponding rate is unavailable, not zero.
Compare competitors without hiding the method
Start by applying the same per-answer citation rule to each competitor domain. This shows which competitors are cited most often within your panel without requiring a blended score.
If you also report share of voice in AI answers, define the inclusion list and counting unit. One workable citation-based method is your domain’s answer-level citation count divided by the sum of those counts for a fixed set of competitor domains plus yours. An answer citing several tracked domains contributes once to each domain’s count. Keep publisher and directory sources visible separately rather than silently treating them as competitors.
If no tracked domain earns a citation, this share-of-voice measure is unavailable, not zero. Do not change the tracked competitor set between periods without labeling the break. A better percentage caused by removing a frequently cited competitor is a reporting change, not a visibility gain.
The proof: #10’s own citation baseline
Our existing nr10.com progress-report tool provides a concrete starting point: it keeps target-query rankings separate from Google AI Overview observations and records the domains cited for a defined set of buyer prompts. The richer evidence log and counting conventions above are recommendations for extending that baseline, not a claim that this tool already captures every field.
In our June-period 2026 report, generated July 1, an AI Overview appeared for all seven buyer prompts tested, and nr10.com was cited in none. The checks used DataForSEO’s live Google AI Overview sample with US, English, and desktop settings. The prompts included “which agencies do generative engine optimization GEO” and “agency for AI search optimization for online stores.” The report also listed the domains cited instead.
That is a dated baseline for seven Google AI Overview checks. It is not evidence that every assistant ignored #10, and it is not a claim about our visibility today. Its value is specificity: it identifies the questions, the surface, whether the answer appeared, and whether our domain was cited. A later check can be compared with that evidence rather than with a recollection of a favorable screenshot.
The operational proof is this prompt-level Overview and cited-domain tracking. We are not presenting an unverified citation lift or revenue result.
Pair citations with traffic and commercial evidence
Our visibility framework separates earned citations, referral traffic, engine-provided reporting, and whether a reporting surface is actually available for the property. Each answers a different question.
Citation checks show what appeared in sampled answers. Referral analytics show visits with an identifiable source. Inquiries and purchases show commercial activity under your attribution rules. None should silently stand in for the others.
A citation with no recorded visit is not automatically worthless. A visit attributed to an assistant does not establish which prompt produced it. Equally, an increase in sampled citations does not prove that an increase in inquiries was caused by AI search. Report the observations together and be explicit about the missing connection.
Where an engine offers webmaster reporting, verify what is accessible for your property and what the metric covers before including it. A feature announcement is not evidence that you have a usable report.
Turn the scorecard into a page-level work queue
A measurement loop is useful when it changes what you do next. We recommend weekly checks for priority prompts, monthly checks for broader category questions, and targeted rechecks after major content updates, consistent with our tracking framework.
Review repeated gaps before commissioning more content. If your brand is described incorrectly, check the facts on owned pages and external profiles. If a competitor’s comparison page is repeatedly cited, inspect the question it answers and the evidence it provides. That is a lead for investigation, not proof of the engine’s selection logic.
Map each proposed change to a page, an owner, and a dated recheck. Protect existing search value: improve established pages additively rather than replacing useful sections to chase one AI answer. Keep the original prompt panel running so a new panel cannot conceal a regression.
This measurement sits within our SEO, AEO and GEO work. The broader method combines discoverability, extractable answers, and verifiable proof. Measurement tells us where to investigate; it does not make a particular schema type or publishing cadence a guarantee of citation.
FAQs about AI visibility
How do we measure share of voice in AI answers?
Choose a fixed buyer-prompt panel, surface, and competitor set. Count brand mentions or domain citations using a declared rule, then compare the same panel over time. Label mention-based and citation-based share of voice separately, retain the raw counts, and disclose panel changes.
Does ChatGPT visibility mean the same thing as a citation?
No. ChatGPT can mention or recommend a brand without linking to its website. Record the mention, recommendation, and owned-site citation separately, along with whether search was enabled and which sources were displayed.
How to get cited in AI Overviews: what should we measure first?
First establish whether an Overview appears for each target question, which domains it cites, and whether it links to yours. Then inspect the cited pages and your candidate page. Better direct answers, clearer facts, internal links, and appropriate schema are work to evaluate, not a promise that Google will select your page.
How should an AEO or GEO report handle missing answers?
Separate collection failures from observed responses. A timeout or access block leaves the result unknown. A successfully observed refusal is a response with no recommendation; retain it with a refusal label. A successful Google search with no AI Overview belongs in the Overview-availability count, not the citation-rate denominator for observed Overviews. Declare these rules before comparing periods.
Can one AI visibility score tell us whether the work is paying off?
Not by itself. Ask which prompts, engines, dates, and counting rules produced it. Use surface-level mention and citation evidence alongside accuracy checks, identifiable referral visits, and commercial outcomes. Keep those signals separate enough to diagnose what changed.
Start with a baseline you can inspect
Before buying another dashboard, choose the buyer questions that matter and decide what a citation will mean in your report. Keep the answers, source links, conditions, and counts. Then repeat the checks and use the gaps to direct page-level work.
If you want #10 to help define that baseline and prioritize the fixes, talk with us about AI visibility. We can start with your priority buyer questions and the pages you expect to earn the citation, rather than a content-volume target.
