Most brands find out they have an AI visibility problem the same way: a prospect mentions they asked ChatGPT for recommendations and a competitor came up instead.
At that point, the usual instinct is to buy a tracking tool, but that isn’t necessarily the right first move. A tracking tool tells you which way the line is going but it doesn’t tell you why the line is where it is, and it can’t tell you anything at all about the months before you installed it. So first, you need a baseline.
This guide walks you through five simple checks you can run to gain insight into the current state of your AI search visibility. It takes an afternoon and a spreadsheet.
What is an AI visibility audit?
An AI visibility audit is a point-in-time assessment of how AI search engines see your brand.
The purpose of the audit is to answer four questions:
- whether AI crawlers can reach your site;
- whether the engines mention you when buyers ask relevant questions;
- whether what they say about you is accurate;
- and which sources they pull that information from.
It’s different from AI visibility tracking. Tracking is ongoing measurement of change over time. An audit is the snapshot everything else gets measured against. You do the audit once, properly, then track.
What do you need before you start?
To start, you need a spreadsheet, access to your robots.txt, and access to your server logs or Cloudflare analytics. You’ll also need accounts on ChatGPT, Perplexity, Google AI Mode, Claude, and Gemini. Free tiers are totally fine.
Then, block out three to four hours and use a clean browser profile or incognito windows throughout. Personalization will quietly contaminate your results otherwise, and you won’t know it happened, so use incognito.
Check 1: Confirm the crawlers can actually reach you
Start here. A blocked crawler explains most of what you’ll find in the next four checks, and it’s the cheapest thing to get wrong. So, open your robots.txt file. The thing to understand is that each AI company runs several separate bots, and the one that matters for search visibility is rarely the one people check.
OpenAI runs GPTBot for model training, OAI-SearchBot for surfacing sites in ChatGPT search, and ChatGPT-User for fetches a person triggers directly. OAI-SearchBot is your priority. Sites opted out of it won’t be shown in ChatGPT search answers, though they can still appear as navigational links. OpenAI notes that robots.txt may not apply to ChatGPT-User, so expect to see it in your logs either way.
Anthropic runs ClaudeBot for training, Claude-SearchBot for search result quality, and Claude-User for person-triggered visits. Claude-SearchBot is the one that matters. If you’ve only ever checked ClaudeBot, you checked the training crawler.
Perplexity runs PerplexityBot, which surfaces and links sites in results and respects robots.txt, and Perplexity-User, which fetches on demand and generally ignores robots.txt because a person asked for it.
Google-Extended is the one that trips up nearly everyone. It isn’t a crawler at all. It’s a robots.txt token, and Googlebot does the actual crawling. It governs whether already-crawled content trains future Gemini models and grounds Gemini Apps and Vertex AI. Google states it doesn’t affect inclusion in Google Search and isn’t a ranking signal. Blocking Google-Extended won’t remove you from AI Overviews or AI Mode, because those are features inside Search running on ordinary Googlebot access. If you want control over how you appear in Google’s AI answers, look at nosnippet, data-nosnippet, and max-snippet instead.
Then go past robots.txt, because robots.txt is a polite request rather than an enforcement mechanism. Pull your server logs or Cloudflare bot analytics and check whether those user agents are actually showing up, and what status codes they’re getting.
This is where the real problems hide. We’ve picked up sites with a completely open robots.txt where a WAF rule or a bot-protection preset was returning 403s to every AI crawler. One client had been blocking PerplexityBot and OAI-SearchBot for four months through a security setting nobody remembered enabling. The robots.txt said yes. The firewall said no.
If you find a block, find the date it went in. That date is usually where your visibility problem starts, and it makes the rest of the audit much easier to interpret.
One thing worth saying plainly, because a lot of GEO content gets this wrong: blocking these crawlers doesn’t make you invisible. Engines still pick up brands through search partners, existing indexes, third-party coverage, and user-triggered fetches. That’s exactly why an opted-out page can still show as a link. What blocking costs you is the chance to be retrieved and cited on your own terms, which is worse but not the same thing.
Check 2: Take your baseline snapshot
Build a list of 20 to 50 prompts your buyers would realistically type. For a Web3 brand that usually breaks down into four groups:
- Category questions: “Best self-custody wallet for Bitcoin,” “top neobanks for European users.”
- Alternatives and comparisons: “[Competitor] alternatives,” “[You] vs [Competitor].”
- Use-case questions: “How do I move Bitcoin off an exchange,” “crypto tax software for high-volume traders.”
- The sensitive ones: Pricing, fees, jurisdiction, security incidents. Include these especially. They’re the questions that decide deals and the ones most audits skip because the answers are uncomfortable.
Run the whole list across ChatGPT, Perplexity, Google AI Mode, Claude, and Gemini. Log four things per prompt: whether you appear, roughly where in the answer, which competitors appear, and which sources get cited.
You can do this in one sitting. AI answers drift day to day, so a snapshot collected over three weeks is noise. If you can only manage half the list today, run half the list and note the date. This is the step that people often skip. Everything you measure later only means something because this exists.
Check 3: Read what the engines actually say about you
Presence and accuracy are different things, and most audits stop at presence.
For every prompt where you appear, check three things. Is the description right? Is the pricing or fee structure right? Is it recommending you for work you actually do?
Wrong answers are more common than you’d expect. Stale funding rounds, discontinued products, fee schedules from two years ago, features you removed, jurisdictions you exited.
Then read for tone. There’s a real difference between an engine listing you as an option and an engine describing you as the safe choice for a specific kind of buyer. The second one is what you’re actually competing for.
For scale on how neglected this is: Fractl surveyed 150 marketers in Q2 2026 and found 27% said their brand had been misrepresented in an AI answer, while only 24% monitor for it at all. Small sample, and Fractl sells into this space, so treat it as directional. But it does match what we see as well. An afternoon on this check usually turns up at least one thing worth fixing that week.
Check 4: Audit the sources
This is where most of the actual work comes from.
For each response, write down every page that got cited. Patterns emerge quickly, and they’re usually not your website. Expect review sites, Quora threads, roundup posts, industry publications, comparison sites, and a surprising amount of Reddit.
Two findings should shape how you read your results. SurfacedBy analyzed roughly 16,400 AI answers to buying and brand questions collected between 29 March and 27 June 2026, and pulled 127,198 citations out of them. Counting at the domain level, 69.6% of cited domains appeared on only one engine, and the engines cite at very different volumes: Gemini averages 11.0 sources per response, Perplexity 8.6, Google AI Mode 7.8, Claude 6.8, and ChatGPT just 3.7.
That’s a commercial query sample over a three-month window, but both numbers point in the same direction. You can’t audit “AI” as a single thing. ChatGPT citing under four sources is a much narrower door than Gemini citing eleven, and the door opens onto different pages.
Whichever domains keep appearing for your category are your off-page roadmap. That’s where digital PR and link-building budget should go, and it’s a far better targeting method than a DR filter.
We’ve watched this work in both directions. For several of our clients, publisher placements we built for traditional SEO reasons started showing up as ChatGPT citations for category questions within a couple of months. Listicle and comparison formats performed noticeably better than straight editorial, which makes sense once you’ve seen how often engines reach for a “top 7” page to answer a recommendation question.
Check 5: Make sure your own story is consistent
Models are working out what you are from scattered evidence across the web. Contradictory evidence makes them hedge or pick wrong.
Pull up your homepage, your LinkedIn company page, Crunchbase, your category’s review site, and your Wikidata entry if you have one. Compare four things across all of them: the one-line description, the category you claim, the founding year, and the headline pricing or fee structure.
You will almost certainly find disagreements. Half-updated profiles are the norm, not the exception, and each one is a small vote for the wrong answer.
Then, check your markup. An Organization or Person schema with accurate sameAs links provides search engines with an explicit, machine-readable statement that all those profiles refer to the same entity.
Being straight about the limits here: we can’t show you proof that schema makes ChatGPT or Gemini consolidate your entity, and anyone selling you that certainty is guessing. What we can say is that it’s an unambiguous signal, it costs about an hour, and markup will never fix contradictory profiles on its own. Fix the profiles first, then add the markup.
This is the least glamorous check on the list, and it’s the first thing we fix for new clients, because it’s cheap and it moves.
What do you do with what you find?
Work in this order. It’s roughly cost-to-fix ascending, and it front-loads the things that unblock everything else.
- Crawler blocks. Do it today. A firewall rule change is free, and nothing else works until it’s done.
- Factual errors about your brand. Do it this week. Correct the profiles that are feeding them, starting with whichever domains showed up most in Check 4.
- Entity consistency. Do it this month. Align the profiles, then add the schema.
- Off-page gaps. Do it this quarter and ongoing. Pursue coverage on the domains your citation audit surfaced, in the formats that got cited.
- On-site content. Ongoing. Now that you know which questions you lose, you know what to write.
How often do you run your AI visibility audit?
Re-run the full audit every six months. Re-run Check 1 monthly, because firewall and CDN settings change without anyone telling marketing, and it takes ten minutes.
Between audits, track a subset. Ten to fifteen of your highest-value prompts, run monthly on the same day, in a clean session. That’s enough to spot a real move without turning measurement into someone’s full-time job, and keep the original baseline.
In six months, when someone asks whether the AI search work paid off, the baseline is the entire reason you can answer.
Getting help with it
Everything above is genuinely doable in-house, and we’d rather you ran it yourself than didn’t run it at all.
Where it gets harder is Check 4. Reading a citation pattern and turning it into an off-page plan that actually earns placements on those domains is a different skill from spotting the pattern, and it’s most of what we do at Rise Up Media.
If you’d like a hand with any of it, or you’ve run the audit and want a second opinion on what it’s telling you, get in touch. We’re happy to look at what you’ve got.
