An AI visibility audit that shows where AI engines name you, and where they do not

First, are you named? For each buyer prompt, does the engine mention your brand at all, and in what position among the options it lists? Second, are you cited? Does the engine show your website as a source, or does it cite a directory, a review site or a competitor instead? Third, is it correct? Does the engine describe your services, location, pricing and credentials accurately?

Those three results can differ. A brand can be named but never linked. It can be cited for an old page that no longer matches the business. It can also be described wrongly with confidence. Collapsing everything into one score hides the useful detail, so we keep them separate. Our guide to brand mentions vs citations explains why.

• Named: does the brand appear in the answer?
• Cited: is your own site shown as a source?
• Correct: are the facts about you right?
• Competing: who appears instead, and which sources support them?

A bad prompt set produces a pretty and useless report. If every prompt includes your brand name, you will look great and learn nothing. If every prompt is a generic head term, you will look invisible even when you are the best local option.

We build the set from your real buyers. We look at sales calls, your search data if you share it, and the questions competitors answer on their sites. We mix stages of the decision: learning, comparing, choosing and verifying. For UAE, USA and UK clients we also set the location and language the way the buyer would. There is more in how to research buyer prompts for GEO.

You approve the final list before testing starts. Some owners want a prompt on the list because a competitor boasts about it. We will include it if it reflects a real buyer question.

Ask each engine the questions your buyers ask, without including your brand name, and record who is named and which sources are cited. Repeat on a different day. If you appear in some answers and not others, you have partial visibility. If you never appear, find out why: access, content, entity facts or missing third party mentions.

That manual test is the core of every audit. What separates a quick look from a proper audit is the discipline around it: a fixed prompt list, the same wording each time, several engines, dated records, and a diagnosis for each gap. You can do the first pass yourself with the free AI Visibility Checker.

This page explains what a full audit covers, how we pick prompts, how we score results without pretending to precision, what the typical causes of invisibility are, and what you can do with the findings.

A prompt set is only as useful as its realism. We build it from the questions your sales team hears, your search query data, your competitors' positioning, and how buyers in your market phrase things. For a UAE business, that can include Arabic and English phrasing and city names. For a US or UK business it includes local terms and spellings.

We group prompts by intent and funnel stage. Problem aware prompts describe a pain. Solution aware prompts compare approaches. Vendor aware prompts name a shortlist. A prompt such as best provider of X in Dubai is a shortlist prompt. A prompt such as what is X is a learning prompt. Both matter, but the shortlist group is where being named has direct commercial value.

We also include a few branded prompts, such as what do you know about your company name, to check accuracy. These are not the main score. They show what an engine believes about you when asked directly. See how to research buyer prompts for GEO for the full method.

Twenty prompts is the Pilot limit, and it is enough to find patterns. It is not a statistical sample of all questions in your market.

• Problem aware, solution aware and vendor aware groups
• Local wording, city names and spellings for your market
• A few branded prompts for accuracy checks
• A fixed list, frozen before the retest

The visibility test tells you what happens. The technical and entity audit tells you why. We check robots.txt and any firewall or CDN rules that affect named AI crawlers, because a blocked crawler cannot read your pages. We check whether the sitemap lists the right canonical pages, and whether key content is present in the HTML or only appears after scripts run. See technical GEO for the crawler list and the reasoning.

We check structured data against the visible page, and whether the Organization data names the business the same way as your other profiles. We check the About page, contact details, leadership names, and the facts a stranger would need to trust you. Then we compare these to what directories, review sites and press say.

The output is a ranked list of blockers, from those that stop engines reading you at all to those that make you slightly less clear. The audit lists them. The Pilot fixes them.

Where you give us access to analytics, we also look at referral traffic from AI engines, so the audit has one more evidence stream beyond the prompt tests. This traffic is often small and mixed into other groups, so we report it as a rough signal only, with the limits stated.

• Robots.txt, firewall and CDN rules for AI crawlers
• Sitemap and canonical consistency
• Content present in raw HTML
• Structured data matches the visible page
• Entity facts match across profiles

Example, hypothetical: a London accountancy firm tests twelve prompts. It is named in two answers in Perplexity, one in Copilot, and none in ChatGPT or Gemini. Sources cited for competitor mentions are mostly a professional directory and two industry blogs. Its own site is cited once.

The audit finds that the firm's service pages describe packages in a PDF only, its company name appears in three slightly different forms across the web, and its Bing indexing is incomplete. The findings are ordered: fix indexing and the PDF first, standardise the entity name, then rewrite the three key service pages with direct answers. The audit also records that the directory is a recurring source and notes it as a future mention target.

None of that is a promise of a result. It is a diagnosis you can act on, with the likely owner for each fix.

Sort the findings into three piles. Fix now: blockers that are cheap and clear, such as an accidental crawler block or a broken sitemap. Fix next: content and entity work that needs writing or approval. Fix later: authority and mention work that takes months. The order is the point. Many teams start with the hardest item and stall.

If you cannot implement the fixes yourself, the report should still be usable by any developer or editor. Ours lists the exact change, the page, the reason and how to verify it. If you want us to make the changes, that is what the GEO Pilot does for $999 USD in 30 days.

If you find wrong claims about your brand, follow how to fix wrong information about your brand in AI. The fix almost always starts with correcting the sources the engine used, not with arguing with the engine.

It can give you a share figure, such as the number of tested prompts where you were named out of the total. That is a fair, simple summary of a defined test. It is not your market share, and it is not a ranking. Treat any single visibility score with care, including ours, because it depends on the prompts chosen, the engines tested, the date and the location.

We report counts and patterns, not a single grade. For each engine we show how many prompts named you, how many cited you, and how many described you correctly. We add the prompts where a competitor was named instead. This keeps attention on what to fix. For wider definitions such as citation share and share of voice, see AI visibility metrics.

If you compare two audits from different providers, compare the prompt lists first. Different prompts produce different results even on the same brand on the same day.

One more caution on sampling. Some engines personalise answers, and some vary by location or by whether you are signed in. We note the conditions of every test so that the retest runs under the same ones. If conditions differ, the comparison is weaker, and the report says so plainly. A small, careful test beats a large, sloppy one.

It cannot show every answer an engine gives to every person. No outside party can see that. It samples. It cannot prove why an engine chose a source, because the engines do not publish their full reasoning. We offer evidence based explanations and mark them as such.

It also cannot predict your result after fixes. A good baseline lets us choose smart changes and then test the effect. It is not a forecast. The details here are accurate as of October 2026, and engines change often, so a baseline older than a few months should be rerun.

• A sample, not a census of AI answers
• Reasons for citations are inferred, not confirmed by the engines
• No forecast of results, no promised citations