A good GEO audit ends with a work plan, not a score. Run this one and you will know, per assistant, which buyer questions your brand loses, which competitor wins each of them, and the page that earned the win. That is a different job from checking whether AI can describe you correctly, and it is the job that changes what you publish next.
If the question you are chasing is “can AI even understand what we do,” start with our brand-understanding audit, which checks whether your public facts are clear and consistent. This guide is the next layer: whether AI recommends you over competitors, and where it does not. It matters because absence is the default. Across about 26,400 AI answers we logged, the brand a question was about was missing from roughly 84 percent of them, so the useful question is not “are we there,” it is “where, and who is there instead.”
Step 1: Build the prompt set from real buyer questions
The audit is only as honest as its questions. Write 15 to 20 prompts a real customer would type, not head-term keywords. Cover the range they actually ask across: the category (“best CRM for a small agency”), comparisons (“X vs Y”), specific use cases, alternatives (“alternatives to X”), and the objection and logistics questions that come up right before someone buys. Include the commercial “best X” questions explicitly; those are where money changes hands.
Keep the list fixed. It is both the audit instrument and the scoreboard, and it only works if you re-run the exact same prompts later to measure change.
Step 2: Run one benchmark, then spot-check
Run every prompt across ChatGPT, Claude, Gemini, Perplexity, and Google AI Mode on a single day. That is your benchmark, the baseline every later comparison anchors to. Then spot-check a subset over the following week or two, because AI answers vary between runs and you want the stable pattern, not one day’s noise.
For each answer, record one of four outcomes for the brand: cited (its page is in the source list or linked inline), named (the brand appears in the answer text with no link), considered (pulled into the retrieval pool but not shown to the reader), or missing. Keep cited and named separate. A brand that gets named with no link is in the conversation and earning none of the click, which is a different problem from being absent, and collapsing the two hides it. For every prompt you lose, also write down which competitors were cited instead, and where their pages sat in the source order.
Step 3: Score per engine, never as one number
Turn the tally into two numbers, computed per assistant: Brand Coverage (the share of answers where the brand was cited or named) and Share of Voice (the brand’s share of all appearances against the named competitors). Do them for each engine separately.
The five assistants disagree sharply about who they will cite for the same question, so a single blended score averages away the surface where you are actually invisible. One engine can name you in a fifth of its answers while another never does, and the fix is not the same on both. We measured how little the engines’ cited webs overlap in the engine-overlap study; the practical consequence for your audit is that the per-engine split is the finding, and the average is the thing that hides it.
Step 4: Split the demand into two shelves
Now line your results up against search demand and sort the questions into two shelves. The informational shelf is the how, where, and when: routes, guides, definitions, comparisons of approaches. The commercial shelf is the “best X to buy” and “who should I hire” questions, where someone is choosing a provider.
A common pattern to look for is a brand that owns the informational shelf, ranking well and getting cited for the how-to questions, but absent from the commercial shelf, where assistants recommend providers instead. If that is your audit, you do not have a visibility problem across the board. You have a specific gap on the questions that convert, and that gap is the plan.
Step 5: Reverse-engineer the pages that beat you
For each lost question, open the competitor page the assistants actually cited and read it as a shape, not as content to copy. Note three things: the question the page answers in its title, how quickly the first concrete fact appears, and where the brand name sits relative to that fact. The winners are almost never better brands. They are better-shaped pages: one focused, factual, retrievable page built to answer exactly that one question.
Two writing rules come straight out of that reading, and they are what make your own replacement pages citable. Lead each answer with a complete, quotable fact-sentence, one an assistant can lift whole, rather than a paragraph that reaches the number in line four. And put the brand name next to the fact, so that when the model reuses your material your name travels with it. A fact sitting in narrative prose with no brand attached gets borrowed and credited to whoever did attach their name to it.
Step 6: Rank the off-site gaps by evidence
Most AI answers are assembled from third-party sources the engines already trust, not from a live read of your own site, so part of the audit is mapping those sources. From your recorded citations, list the review sites, directories, roundups, community threads, and video pages that keep getting cited on your prompts, and mark whether your brand appeared on each.
Rank that list by the evidence you just collected, how often a source was cited and whether you were on it, not by generic domain authority. A review site an assistant quoted by name when it decided how strongly to recommend a category is worth more than a high-authority site that never comes up for your questions. That ranking is your off-site priority order.
Step 7: Turn it into a 90-day plan with a scoreboard
Order everything by leverage. Reputation and structured-data fixes that take an afternoon go first, the pages that close the biggest commercial gaps go next, and the slower off-site and content work fills the 60-to-90-day window. Every item should trace back to a specific lost prompt, so nothing on the list is there for its own sake.
Then set the scoreboard and name the limit in the same breath. The scoreboard is the re-test: run the identical prompt set across the same five assistants at day 90 and compare against the benchmark. The limit is that no one can guarantee an assistant will recommend a brand, and an audit that says so plainly is the one worth trusting. What the plan changes is the surface area of quotable, retrievable, brand-attached facts and third-party proof, which is the part of the system you control.
One last discipline: do not chase every lost prompt with a thin page. Five focused pages written to your standard beat fifteen rushed ones, and thin pages built only to catch an assistant tend to read that way to readers too. If you are scoping the measurement side of this, our rundown of AI SEO tools compared covers what a report needs to show, cited sources and per-engine coverage, not just a mention count.
