Start with the free AI Visibility Score. It checks your site against the first layer of this audit in about a minute and shows the result without asking for an email. Then use the rest of this guide to run the other two layers properly, or to sense-check what the tool found.
An AI visibility audit answers one question: when a buyer asks an assistant about your category, does your brand appear, and if it does not, who appears instead?
Most businesses have never run one, because the thing it measures did not exist as a channel three years ago and no standard SEO tool reports it. The method below is the one we use, split into three layers, with the scoring and the fix prioritisation attached.
The three-layer AI visibility audit
A complete audit covers three layers, and they have to be done in order, because each one explains the results of the one before it.
| Layer | What it measures | Roughly how long |
|---|---|---|
| 1. GEO visibility | How often and how accurately your brand is cited in AI-generated answers | Half a day for 20 questions |
| 2. AEO presence | Which snippets, rich results and structured data your pages hold in Google | Half a day |
| 3. SXO alignment | Whether the traffic layers 1 and 2 produce lands on pages that convert it | Half a day |
Skipping straight to layer two is the common mistake. Schema is easy to audit and produces a satisfying list of fixes, which makes it tempting to start there. It tells you nothing about whether you are being cited.
Layer 1: GEO visibility audit
This layer is manual. There is no report that shows it, because a citation inside a generated answer is not a ranking position.
Build the question set
Write 20 questions an assistant should be answering in your category. The discipline that matters most: do not search your own brand name. A model handed your name will usually find you, which tells you nothing. You are testing whether you are the answer to a question that does not mention you.
Weight the set towards buying intent:
- Category questions. “Who are the best [service] firms in [city]?”
- Comparison questions. “Should I use [approach A] or [approach B]?”
- Qualification questions. “What should I look for when choosing a [service] provider?”
- Problem questions. The symptom a buyer describes before they know what to buy.
- Price questions. “What does [service] typically cost in the UK?”
A set made only of general category questions produces a flattering result. The specific commercial questions are where most brands disappear, and those are the ones with money behind them.
Run each question across the platforms
For each question, run it in ChatGPT, in Perplexity, in Google (checking whether an AI Overview triggers), and in Gemini and Claude if your buyers use them. Use a fresh session with no chat history, because personalisation from your own previous prompts will quietly flatter you.
Run each question at least twice. Assistants are not deterministic. The same question can return a different set of brands on consecutive runs, and a brand that appears once in three runs is in a genuinely different position to one that appears every time.
For each run, record four things:
- Does your brand appear at all?
- If it appears, is it named first, or fourth in a list?
- Is the description of you accurate?
- Which sources did the answer cite, and what do those pages have in common?
The fourth is the one people skip and the one that pays. The cited sources are a list of exactly what the model considers a credible answer to that question. If the same three domains appear across fifteen of your twenty questions, you have found your real competitive set, and it is often not the companies you thought you were competing with.
Score it
Two points for a direct citation where you are named. One point for a mention without attribution. Zero for no presence. Twenty questions gives a score out of 40.
| Score | What it means |
|---|---|
| Below 15 | Significant gaps. The brand is largely absent from answers in its own category. |
| 15 to 25 | The most common result. You appear on general questions and lose the specific commercial ones. |
| 26 to 34 | A real position, usually with one or two platforms much weaker than the rest. |
| 35+ | Strong. The work shifts from earning citation to defending it. |
Score each platform separately as well as in total. A brand strong in Perplexity and invisible in ChatGPT has a structural problem, not a content one, and the fix is different.
Layer 2: AEO presence audit
This layer uses tools, and most of them you already have.
Featured snippets. Export your Search Console queries at positions one to ten and check which return a featured snippet. Note which you own and which a competitor owns. A page ranking third on a query whose snippet someone else holds is losing the answer without losing the ranking.
Rich results. Run your top ten pages through Google’s Rich Results Test. Record which have valid FAQ, Article, Service or Organization markup and which are throwing errors. Errors matter more than absences here, because an invalid block can suppress a result that would otherwise show.
Structured data coverage. Crawl the site and audit schema by page type rather than page by page.
The question is not “does this page have schema” but “does every page of this type have the right
schema”. Service pages without Service, articles without Article, question pages without
QAPage.
Entity consistency. Check that your organisation name, address and description are identical across your site, your Google Business Profile, Companies House and your main directory listings. Models corroborate across sources, and contradictions between them suppress confidence. This is the cheapest fix in the entire audit and one of the most effective.
Crawler access. Confirm your robots.txt is not blocking the AI crawlers you want retrieving
you, and that key pages are not rendering their content only in JavaScript.
Layer 3: SXO alignment audit
Visibility that does not convert is a vanity result. This layer checks whether the first two are producing anything.
Intent match. For your top ten landing pages by organic traffic, compare the main query driving them against what the page actually offers. Rate each as strong, weak or mismatched. A mismatched page is usually a page ranking for a question it does not answer.
Enquiries, not sessions. Compare conversion rate for organic against direct. Organic underperforming badly usually means intent mismatch rather than a traffic problem. Count enquiries, calls and form submissions rather than sessions, because sessions can rise while enquiries stay flat and that looks like success on a chart.
Zero-click reality. Some of your best pages will show rising impressions and flat clicks. That is not automatically failure. It can mean an AI Overview is answering in place while still citing you. Check whether you are the cited source before treating it as a loss. The zero-click statistics page has the detail.
Page experience. Run the homepage and top five pages through PageSpeed Insights and record LCP, INP and CLS. This is last for a reason. It matters, and it is almost never the thing holding back citation.
If the audit points towards bringing help in, the questions worth asking an AI SEO agency will save you a wasted first call.
Building the fix roadmap
Sort every gap into four boxes, then work them in this order.
High impact, low effort. Entity consistency corrections, schema fixes, heading reformatting to target snippets, intent-match corrections on high-traffic pages. This box is usually where the first month of measurable movement comes from.
High impact, high effort. Answer-first rewrites of the pages losing your most commercial questions, new pages for question clusters where nothing of yours exists, off-site corroboration on the domains your layer one audit found being cited.
Low impact, low effort. Minor schema additions, FAQ expansions, image optimisation. Useful for filling a schedule, not for making the case that the programme is working.
Low impact, high effort. Leave it. Core Web Vitals work on a page nobody is citing belongs here more often than people expect.
Re-run layers one and three quarterly against the same question set. A single audit gives you a number. The same audit repeated gives you a trend, and the trend is the only thing that tells you whether the work is doing anything.
For the method behind all of this, the Synaptic Authority Engine sets out the full six-step process. If you would rather commission the audit than run it, our AI search audit pricing is published, and the free AI visibility audit is the no-cost version of layer one.
AI visibility audits: common questions
What is an AI search visibility audit?
It is a structured check of whether AI assistants cite your brand when someone asks about your category, which structured results your pages hold in Google, and whether the traffic either one produces lands somewhere that converts. It covers three layers: GEO visibility (citation inside AI answers), AEO presence (snippets, rich results and schema) and SXO alignment (whether the visibility turns into enquiries). Running it manually takes most teams two to three days.
How do I check if ChatGPT mentions my business?
Ask it directly, using the questions a buyer would ask rather than your brand name. Searching your own brand name tells you almost nothing, because a model will usually find you if it is handed the answer. The useful test is a category question such as “who are the best commercial law firms in Manchester”, run without personalisation and repeated a few times, because answers vary between runs. Record whether you appear, whether the description is accurate, and which brands appear instead.
How many queries should an AI visibility audit test?
Twenty is the practical floor for a self-run audit and gives a usable score out of 40. Fewer than that and one unusual answer distorts the result. MarGen’s own Synaptic Audit uses 15 commercial queries across five platforms, which is 75 data points, because platform variance matters as much as query variance.
Should I test the same question more than once?
Yes. Assistants are not deterministic, so the same question can return different brands on consecutive runs. Run each question at least twice, in a fresh session with no chat history, and record both. A brand cited once in three runs has a real but weak position, which is a different problem to one that never appears.
Does ranking in Google still matter for AI citation?
Yes, and it is the single strongest lever most businesses have. Assistants retrieve from the open web, and the pages they retrieve are disproportionately pages already ranking. That is why the audit checks Google structured results as its second layer rather than treating AI and traditional search as separate problems.
What score means I have a problem?
Below 15 out of 40 on the layer one scoring means significant gaps, where the brand is largely absent from answers in its own category. Between 15 and 25 usually means you appear on general questions and lose the specific commercial ones, which is the most common pattern and the most expensive, because the specific questions are the ones with buying intent behind them.
How often should the audit be repeated?
Quarterly for the full three layers. Model versions change, retrieval behaviour changes, and competitors publish. A baseline that is nine months old tells you very little about where you stand now, which is why a single one-off audit is worth much less than a repeated one measured against itself.
Can I audit AI visibility with normal SEO tools?
Partly. Search Console, the Rich Results Test and a crawler cover layer two well. Layer one has no equivalent, because citation inside a generated answer is not a ranking position and does not appear in any standard report. That part is either manual, done against a fixed question set, or handled by a tool built for it.
Where to go next
Related: for the off-site half of the audit, meaning every public source describing your business and where they contradict each other, see how to audit your off-site AI ecosystem. For what to do with the findings, see AI visibility strategy.