AI visibility

Definition: an outcome, measured over a panel

Geo-visibility is the degree to which a brand appears in the answers that generative systems give to a defined set of questions. It is measured, not practised. The output is a rate with a method attached: on these forty prompts, in this locale, on these dates, the brand was mentioned in 46 percent of answers and cited as a source in 21 percent.

The term is often used interchangeably with AI visibility, and the glossary entry under that name describes the same idea. What matters is the discipline behind the number. An answer engine composes a response from retrieved sources, so visibility is a joint outcome of being retrievable, being quotable and being corroborated. None of those is directly observable from the outside, which is why the measurement has to be designed rather than scraped casually.

This entry is the measurement companion to the GEO entry at /glossary/geo/. That page covers entity work, page structure, crawl access and corroboration. This page covers how to find out whether any of it changed what a machine says about you. Keeping the two separate is not pedantry; it is the difference between a report that explains a movement and one that merely records activity.

The measures, and what each one can tell you

Mention rate is the share of answers in which the brand appears at all, with or without a link. Citation rate is the share in which the answer links to your domain or names it as a source. Share of voice is your mentions divided by all brand mentions in the panel, which shows relative presence. Citation share is the same arithmetic applied to citations, and it tends to be more concentrated than mentions because engines cite a small set of sources.

Answer accuracy is the qualitative measure that stops the others from being misleading. Being mentioned in every answer is worthless if the answer says you charge a flat annual fee when you charge monthly, or that you operate only in one city. Accuracy is scored per answer on whether the stated facts about you are correct, and it is the measure most likely to reveal a real problem.

Two secondary measures are useful. Prominence asks where in the answer the brand appears, since a citation in a footnote is not the same as being named in the first sentence. Sentiment, or framing, asks whether the description is positive, neutral or critical. Both are softer measures and should be reported as distributions rather than as a single score, because a model’s framing of a brand is unstable across runs.

Designing the prompt panel

A prompt panel is the fixed set of questions you ask repeatedly. It should mirror real buying and research behaviour rather than flatter the brand. Cluster the prompts: category questions such as which CRM suits a mid-market Indian company, comparison questions such as one product against another, pricing questions, implementation and migration questions, and problem questions where the buyer has not chosen a category yet.

Exclude brand-name prompts from the headline number. Asking which company named LoopNow offers an AI agent is a different measurement from asking which agent platform a fifty-person Indian firm should evaluate, and mixing them produces a number that looks good and means little. Report brand prompts separately as a reputation check, because they do reveal whether a model describes you correctly once it knows who you are.

Wording sensitivity is real. Adding the word India, naming a city, specifying team size or mentioning a budget changes retrieval and therefore the answer. Keep the wording stable across periods so the trend is comparable, and run a small wording-variant set alongside to understand how fragile a result is. If a one-word change moves the mention rate by thirty points, the headline number is a statement about wording, not about the market.

• Thirty to fifty prompts, clustered by intent, reviewed quarterly.
• No brand-name prompts in the headline visibility rate.
• Fixed locale, language and account state for comparability.
• A small variant set to test wording sensitivity.
• Store the exact prompt text with every capture.

Sampling, variance and the trap of weekly numbers

Generative answers are not deterministic. The same prompt can produce a different answer on a different day, with a different retrieval set, and with different brands named. Some of that is model versioning, some is index freshness, some is sampling, and some is personalisation. Treating one run per prompt per week as a precise measurement produces dramatic movements that are mostly noise.

Reduce variance with design rather than with hope. Run each prompt several times per period and report the average or the range. Keep the engine, locale and account state fixed. Capture the date and, where visible, the model name. Record whether the answer cited sources at all, because an answer with no citations is not evidence about retrieval.

Then set reporting cadence to match the noise. Monthly is realistic for a panel of forty prompts with three runs each. Weekly is useful only as an operational check on a specific change, and it should never become the number a leadership team reviews, because the first question will be why the line moved and the honest answer will be that it did not, meaningfully.

Collection: manual, tooled, and the personalisation problem

Collection can be manual, tooled or a hybrid. Manual capture is slower and more honest: a person opens a logged-out session, asks the prompt, and saves the answer and the citation list. Tooled capture scales and standardises, but it depends on the vendor’s access method, and a tool that queries through an API may see a different answer from the consumer interface, because retrieval, personalisation and product surfaces differ.

Personalisation is the hard part. A logged-in account in one city, with a search history and a set of connected apps, can receive a materially different answer from a clean session. For a market-level measure, use logged-out sessions and a fixed region. For a customer-level question, such as what an existing user sees when they ask about support, a logged-in capture is the right instrument. State which one you used, because mixing them produces comparisons that cannot be defended.

Save the evidence, not only the score. A screenshot or a saved transcript with the date, the engine and the prompt is what settles an argument six months later. Coding without evidence is an opinion with a spreadsheet, and in a category where answers change monthly, the archive is the asset.

Coding answers without fooling yourself

Coding turns an answer into structured facts: was the brand mentioned, was the domain cited, was the description accurate, which competitors appeared, and which sources were cited. The rules need writing before the coding starts. Does a mention inside a citation title count as a mention? Does a link to a press release on a third-party site count as a citation? Is a generic reference to the category counted, or only a named brand?

Accuracy coding is the most valuable and the most neglected. Build a fact sheet for your own brand: pricing, GST treatment, billing unit, support hours, regions served, integrations, and the things you explicitly do not do. Then score each answer against it. This catches the failure mode that mention rate hides, where a model confidently states a policy you changed a year ago and the report shows improving visibility.

Use two coders on a sample and measure agreement. If the two disagree on a fifth of cases, the coding rules are ambiguous, and the number derived from them is not comparable across months. This is the same calibration discipline an eval programme needs, applied to a smaller and simpler judgement.

A worked baseline, with arithmetic you can check

Take forty prompts, three runs each, on one engine in one locale. That is 120 answers. Suppose the brand is mentioned in 55 and cited in 24. Mention rate is 55 divided by 120, or 46 percent. Citation rate is 24 divided by 120, or 20 percent. If all brands together are mentioned 190 times across those answers, including you, share of voice is 55 divided by 190, or 29 percent. Every number has a denominator you can name, which is the test of a usable metric.

Now slice it. Suppose the 120 answers split across clusters, with category prompts producing 30 mentions out of 45 answers, pricing 12 out of 30, comparison 9 out of 24, integration 3 out of 12 and migration 1 out of 9. The five counts add up to the 55 mentions and the five answer counts add up to 120. The headline 46 percent is being carried by a cluster where you are already strong, and migration, which is where buyers are closest to a decision, is nearly invisible. That pattern is the normal result of a first baseline, and it is more actionable than any single score.

Finally, check accuracy on the answers where you appear. If a third of them describe the pricing model incorrectly, the priority is not more visibility, it is correcting the pages and sources that the model is drawing from. Visibility without accuracy is a brand risk that reports as a win.

Benchmarks, competitors and the honesty problem

Competitive visibility is useful when the method is disclosed and dangerous when it is not. A league table that mixes engines, locales and prompt sets, collected on different days by different methods, is not a comparison; it is a set of numbers arranged in a column. If you publish or present one, state the panel size, the engines, the locale, the dates and whether prompts included brand names.

Internal comparison is more reliable. The same panel, the same engine set, the same coding rules, measured monthly, gives you a defensible share-of-voice trend against named competitors. Even then, treat small differences as ties. A two-point gap on 120 answers is inside the noise, and presenting it as a ranking invites a decision that will be reversed next month.

Watch for entity confusion too. Models sometimes attribute a competitor’s feature to you, merge two companies with similar names, or describe an older product line as current. These are not visibility wins even when they mention you. Coding them separately, as inaccurate mentions, keeps the report honest and gives the team a specific correction to make.

The mistakes that make the number meaningless

The first is a panel made of vanity prompts. Questions that already include the brand, or that are so specific only you could answer them, produce high visibility and teach nothing. The second is prompt drift: the panel is quietly edited between periods, so the trend compares different questions. Version the panel and note every change.

The third is a single engine. Different assistants retrieve from different indexes, cite different sources and personalise differently. Reporting one engine as AI visibility is like reporting one keyword as SEO. The fourth is no accuracy measure, which lets a rising mention rate conceal a wrong description. The fifth is no archive, so nobody can check what the answer actually said when the number was recorded.

The sixth is attribution. Visibility can move because of your work, a competitor’s outage, a model update, a seasonal shift in what people ask, or a change in the retrieval index. Without a practice log recording what changed and when, every movement becomes a story told after the fact. Record the changes, keep the panel stable, and be willing to say that a movement has no confirmed cause.

Cadence, ownership and how it connects to revenue

Run the panel monthly, review it with the people who own content, and publish a short report: panel version, engines, dates, mention rate, citation rate, accuracy, and the weakest cluster with a named owner. Quarterly, review the panel itself, retire prompts that no longer reflect buyer questions, and add new ones from sales calls and support tickets.

Connecting visibility to revenue is indirect and should be described as such. A buyer who sees a brand in an AI answer may search for it, visit directly, or ask a colleague, and none of those paths carries an AI referral. What you can do is instrument the downstream: brand search volume, direct traffic to the pages cited, assisted conversions where analytics can see a referrer, and win-loss notes that mention how the buyer found you.

Do not overclaim. A dashboard that says AI visibility drove a specific amount of pipeline is usually inventing precision. A dashboard that says visibility improved on these clusters, brand search rose in the same period, and sales heard the brand mentioned in three late-stage calls is honest, useful and enough to justify the work.