How to Run a GEO Audit: A 7-Step Checklist
A GEO audit is a structured check of whether AI answer engines (ChatGPT, Perplexity, Gemini, Claude, and Google's AI Overviews), can currently find, understand, and cite your site when someone asks a question your business should own. It is not a technical SEO audit with a new label: a technical SEO audit tells you whether Googlebot can crawl and rank your pages; a GEO audit tells you whether an AI model would quote you in a synthesized answer, which is a different question with different failure modes.
The output of a real GEO audit is three things: a baseline citation rate per engine and prompt, a prioritized list of what's blocking citation, and a fixed cadence for re-measuring. None of it involves guessing. This guide walks through the seven steps in order, start to finish, most audits take two to four hours the first time and under an hour on repeat runs.
Why a GEO Audit Isn't Just an SEO Audit With New Labels
Search engines and AI answer engines fail differently, so they need different diagnostics. A page can rank #1 on Google and never get cited by ChatGPT, and a page that ranks nowhere on page one can still get quoted inside an AI Overview. The two systems read different signals: Googlebot crawls and indexes for ranking; GPTBot, PerplexityBot, ClaudeBot, and Google-Extended crawl (when they're allowed to) to train or retrieve content that a model can later synthesize into an answer. A GEO audit has to check both layers is the content reachable by these AI crawlers at all, and once reachable, is it structured in a way a model can lift and attribute correctly.
This distinction matters because most "SEO audit" checklists stop at crawlability and call it done. A GEO audit goes further: it asks whether the page answers a question directly enough, early enough, and with enough surrounding topical corroboration that a model would choose to cite it over five other pages saying something similar.
Step 1: Establish Your Baseline Citation Rate Before You Touch Anything
Skipping this step is the single most common mistake. Before changing a page, run the actual questions your buyers ask through the engines that matter for your category, and record, honestly, whether and how you're cited today. Without a baseline, you have no way to know later whether any change worked or whether the engines simply answered differently on a different day, which they do.
This is also the fastest step to execute, and it's literally what an AI visibility report (ours included) does: it runs a set of buying-intent prompts against multiple engines and shows you whether and where you're currently cited. You don't need a scan to do this step manually, you can run the prompts yourself in each engine's chat interface and log the results in a spreadsheet, but doing it by hand across five engines and even a modest prompt set gets tedious fast, which is exactly why this step gets skipped in practice.
Step 2: Map the Real Prompts Your Buyers Actually Use
A baseline is only as good as the prompts behind it. "Best [category] tools" is a fine starting prompt, but real buyers ask more specific, more varied questions: comparisons against named competitors, "how do I check whether X is a good fit for a small team," "is there a free way to test this before I commit." Generic prompts tell you whether you show up for the obvious query; specific prompts tell you whether you show up where the actual buying decision happens.
Building this prompt set properly (how many prompts, how to phrase them, how often to re-run them) is its own discipline, covered in detail in our guide to prompt monitoring. For the audit, ten to twenty prompts covering informational, comparison, and "is there a free tool" intents is enough to get a meaningful baseline without turning the audit into a research project.
Step 3: Check Whether AI Crawlers Can Even Reach Your Content
Before worrying about content quality, confirm the content is reachable. Check robots.txt for explicit blocks on GPTBot, ChatGPT-User, PerplexityBot, ClaudeBot, Google-Extended, and any other AI crawler user agents you can identify. Some sites block these by default via a security plugin or CDN setting without anyone deciding to. Confirm key pages aren't gated behind a login wall, aren't rendered exclusively via client-side JavaScript with no server-rendered fallback (a real barrier for crawlers that don't execute JS the way a browser does), and haven't picked up an accidental noindex tag during a redesign.
This is also the point to check for (and, if absent, add) an llms.txt file at your domain root: a plain-text index that tells AI crawlers what your site is and which pages matter most, the same underlying idea as robots.txt and sitemap.xml but purpose-built for the AI-crawler era. It doesn't force a citation, but it removes ambiguity about what you consider your most important content.
Step 4: Audit Content Structure for "Answer Shape"
Once you know a page is reachable, check whether it's readable the way a model needs it to be. Content that states a direct question and gives a direct, self-contained answer in the first one or two paragraphs is far easier for a model to lift and attribute correctly than a page that opens with three paragraphs of scene-setting before it ever answers anything. Look for:
- A clear question-and-answer near the top: not buried under a long narrative introduction.
- Descriptive subheadings that a model (and a human skimming) can use to jump straight to the relevant section.
- Self-contained paragraphs that make sense pulled out of context, because that's often exactly how a model uses them.
- An FAQ section, where it fits naturally, since FAQ-formatted content maps directly onto how many answer engines retrieve and quote.
Pages that read like marketing copy (long on adjectives, short on a plainly stated claim) tend to be the ones that get skipped over in favor of a competitor's more direct answer, even when the marketing copy is technically more persuasive to a human reader.
Step 5: Check Your Structured Data
Schema.org markup (Organization, Article, FAQPage, HowTo) doesn't guarantee a citation, but it gives an engine an unambiguous, machine-readable version of the same information a human sees on the page. Confirm your key pages carry accurate FAQPage markup where you have an FAQ, Organization markup with your name and logo consistently declared site-wide, and that none of it is stale, mismatched structured data (a schema block that says one thing while the visible page says another) is arguably worse than having none, because it signals inconsistency rather than absence.
Step 6: Check Off-Site Authority and Corroboration Signals
AI engines appear to weight pages that are corroborated by other credible sources covering the same topic more heavily than a single isolated page making the same claim with no supporting context. This step is less mechanical than the others: look at who else publishes on your topic, whether your claims are consistent with what's said elsewhere (inconsistency reads as untrustworthy to both humans and models), and whether you have any citations, mentions, or backlinks from sites a model would already consider credible in your category.
This step is slower to move than the others (it's built over months, not fixed in an afternoon), but it belongs in the audit because it explains a lot of the citation gap that structural fixes alone won't close.
Step 7: Re-Measure on a Fixed Cadence and Track the Delta
Run the same prompt set again on a schedule, monthly is a reasonable default, since engines update frequently and a single run is a snapshot, not a stable average. Compare against your Step 1 baseline. This is the only part of the process that actually tells you whether anything worked: not a promise made in advance, but the measured change in citation rate after the fact.
What a GEO Audit Won't Tell You
Be honest with yourself about the limits here. No engine publishes its exact citation logic, and it changes without notice. A single audit run is a snapshot; run-to-run variance is real, which is why the re-measurement cadence in Step 7 matters more than any single number. And no vendor, including us, can promise a specific citation rate as an outcome of fixing any of the above. What a GEO audit gives you is a measured, repeatable process for finding out where you actually stand and what's plausibly in your control to change, which is a meaningfully different thing than a guarantee.
Turning the Audit Into a Priority List
Once you've run all seven steps, you'll typically have more findings than you can act on at once. A reasonable way to sequence them: fix anything in Step 3 (crawlability) first, since a blocked crawler makes every other fix moot. Then Step 4 and Step 5 (structure and schema), since both are one-time, page-level changes with no ongoing cost. Step 6 (authority) is the long game, start it, but don't expect it to move the Step 7 number quickly. And if your Step 1 baseline showed you're already cited reasonably often on some prompts and not others, look for what's different about the pages that are getting cited before assuming the problem is structural at all.
This is the same seven-step process GeoHero's Research Team used to build our own 240-response citation leaderboard, and for how Google's AI Overviews specifically decide what to cite, see our companion guide on optimizing for AI Overviews. If you want Step 1 done for you instead of by hand, get the full report.