Generative Engine Optimization
AI visibility audit
This is the narrowest useful unit of Generative Engine Optimization: find out whether the assistants your buyers use will name you, and why they currently do not.
AI visibility audit: what it is and what it is not
Find out whether the assistants your buyers use will name you, and why they currently do not. In practice that means prompt-set testing across the major assistants for your category, run repeatedly rather than once, and then the harder half: extraction of which sources each assistant actually cites for those prompts.
Why this work exists
Almost every engagement that stalls on this discipline stalls here, and almost nobody predicts it.
It belongs to Generative Engine Optimization and is described here because it is the piece clients recognise and ask for.
The nearest piece of work to it is making the brand the one an assistant names when somebody asks for a recommendation. Where that is already handled, this moves faster than the estimate; where it is not, the two get done together or neither of them holds.
What good looks like
You can tell it is right when a new person can pick it up and be useful inside a week.
The gap between that and where most accounts start is why this is scoped as AI visibility audit rather than as a line inside something larger. Getting the facts about pricing and coverage right in the sources models actually read is the usual second step once it is in place.
When this is the right work
What tells you to start here:
- Your category gets answered by an assistant and your name never appears
- An assistant describes your pricing, coverage or services wrongly
- Referral traffic from assistant domains is zero while competitors see it
If none of those is true, something else is probably the constraint, and the Generative Engine Optimization overview maps the alternatives. The usual neighbours are restructuring the pages so a model can lift a correct answer out of them and checking that the retrieval crawlers are not being turned away at the edge.
What the work covers
What the work actually involves:
- Prompt-set testing across the major assistants for your category, run repeatedly rather than once. The effort here scales with the size of the estate, not with the size of the brief.
- Extraction of which sources each assistant actually cites for those prompts, the step where disagreements surface, which is why it goes early rather than late.
- Gap analysis between what the assistants say about you and what is true. Where the documentation and the actual behaviour part company.
- Crawler access verification for the retrieval bots, in logs rather than in robots.txt alone, with the evidence kept so the decision can be defended later.
How it runs
It runs in three passes. The order is not interchangeable.
Writing it down for the next person
The reasoning gets recorded alongside the result, so a decision taken in March can still be defended in September by somebody who was not there.
Deciding what not to do
A list of things deliberately left alone, with the reason for each. It saves the same conversation happening four times over the following year.
The second look
A fortnight after the first pass, the same ground gets covered again with fresh eyes. The second pass finds a different class of problem from the first.
What each part runs into
The four parts do not fail in the same way:
- Prompt-set testing across the major assistants for your category, run repeatedly rather than once – in Jaipur, complicated by the fact that a market where buyers qualify by area rather than by city
- Extraction of which sources each assistant actually cites for those prompts, which in Gurugram runs into a competitive set that changes street by street
- Gap analysis between what the assistants say about you and what is true: the Hyderabad version of this has to deal with a decision made by somebody who never visits the site
- Crawler access verification for the retrieval bots, in logs rather than in robots.txt alone, where a Indore engagement contends with a procurement step nobody accounted for
The avoidable version of this
Testing one prompt once. Assistant answers vary between runs and between accounts, so a single sample tells you nothing you can act on.
It is worth stating because it is the version of this work that gets sold most often, and it fails in a way that looks like the discipline not working rather than the execution being wrong.
How it is measured
Share of tested prompts where the brand is named, tracked over time rather than as a single figure. That measure is agreed before the work starts, so there is no argument about the result afterwards.
It is worth tracking alongside one commercial number, so a technical win that sells nothing is visible as such.
What you end up with
The deliverable, stated plainly, is a prompt-by-prompt record of who is cited for your category, which of your pages are retrievable, and the specific reasons you are absent.
Structured so one section can be lifted out and sent on without the rest.
How it differs by sector
The sector decides what the constraint actually is:
- SaaS and technology: Product-led companies where the buyer self-educates and the trial is the sales call. so this work has to account for that before anything else.
- Home and trade services, where emergency intent, service-area geography, and a phone that has to be answered. which moves this work earlier in the plan.
- In Travel and hospitality the binding constraint is that aggregators hold the demand, so direct booking depends on being the specific named property.
The industry pages carry the rest of that detail. SaaS and technology and Home and trade services are the two where this piece of work most often decides the outcome.
Common questions
How often should AI visibility be re-tested?
Monthly at minimum, and weekly for a contested category. The same prompt can return different sources on different days as the retrieval index refreshes and as models are updated, so a point-in-time reading is noise. What matters is the trend in the share of prompts where you are named.
Can you be cited by an assistant without ranking in search?
Yes, and it happens often. Several assistants retrieve from their own indexes and from licensed sources, so a page that ranks poorly in classic search can still be the passage a model quotes. The reverse is also true, which is why the two need measuring separately.
How do we know the finding is real?
Because the measurement is shown rather than summarised. Share of tested prompts where the brand is named, tracked over time rather than as a single figure is reported against a baseline taken before anything changed, with the underlying data available. How do we measure SEO for a long sales cycle? deals with the first ninety days, Do we need separate pages for every city we serve? with everything after them.
What comes after this?
Whatever Share of tested prompts where the brand is named, tracked over time rather than as a single figure shows. The next piece is scoped from the result rather than agreed in advance, because committing to it now would be guessing. Conversion tracking is the shortest useful page on this. Cookie consent is the most complete.
Does this work at our size?
The diagnosis does, at any size. Whether the work that follows is worth the money depends on what an improvement in share of tested prompts where the brand is named, tracked over time rather than as a single figure is actually worth to the business, which is the first thing to settle. Marketplace Marketing for Automotive explains the signal. Paid Social and Meta Ads for Automotive explains the noise around it.
Ask about AI visibility audit
Give us the site and one question. If the honest answer is that you do not need this, we will say so.
Have the awkward questions answered first
We would rather tell you the work is not worth the money than take it and report on activity.