AI Search Audit
A fixed-scope diagnostic of how AI systems find, read and describe you
Four weeks, a fixed price and a defined deliverable. We test a prompt panel across the major assistants, check what AI crawlers can actually reach, review structured data against what the page visibly says, and hand over a prioritised list with the evidence behind it.
The audit
Four weeks, in a fixed order
The sequence is deliberate. Access is tested before content, because there is no point assessing the quality of a page that no AI crawler can reach.
Week one — scope and prompt panel
We agree the competitor set, the buyer questions that matter, and the assistants worth testing for your market. From that we build a panel of thirty to sixty prompts, each with a written reason for inclusion, then freeze it so every future run is comparable.
You get: A frozen prompt panel with rationale, and the run protocol
Week one — crawler access testing
Every relevant user agent against every important section of the site: the search crawlers, plus GPTBot, PerplexityBot, ClaudeBot, Google-Extended and the user-triggered fetchers. We record the response each one receives rather than reading robots.txt and assuming the rest.
You get: A crawler access matrix with recorded response codes
Week two — run the panel
Each prompt is run multiple times in clean sessions, with personalisation and memory disabled where the interface allows, and the model version, region and date recorded. Every response is coded for whether you were cited, named, described inaccurately, or absent while a competitor was named.
You get: A coded response set with unedited transcripts
Week three — how you are represented
Every place a machine could find a description of your business, and every disagreement between them. Company name variants, service descriptions, locations, ownership, discontinued offerings still being advertised on somebody else’s site. Wrong claims found in week two are traced back to their sources here.
You get: An entity consistency table with sources and owners
Week three — extractability and markup
Whether the passages on your key pages survive being lifted out of the page, whether claims carry a date and a source, and whether the structured data agrees with what the page visibly shows. Anything marked up that the page does not support is flagged for removal rather than expansion.
You get: A page-level findings list with the specific edit for each
Week four — prioritise and hand over
Findings are sequenced by what unblocks the rest, with effort marked against each item and a clear note on which ones your own team can do without an agency. We walk the report through with you rather than emailing a PDF, and you keep the panel, the protocol and the raw data.
You get: A prioritised backlog, a one-page summary and the full appendix
What you receive
Everything that is in the deliverable
Published in full so you can compare it against anything else you are being quoted for. If a competing proposal will not list its contents at this level, that is informative on its own.
- The written methodology: prompt panel, run count, models and versions, region and dates
- The full response transcripts, unedited, so you can read what was actually said rather than a summary of it
- A coded results table — cited, named, described inaccurately, or absent — with accuracy and sentiment against each response
- Competitor share across the identical panel, from the same runs, so the comparison is like for like
- Run-to-run variation for each prompt, which is the threshold any future change has to beat to count
- A crawler access matrix: every relevant user agent against every important section, with recorded response codes
- Render checks on the templates that carry commercial content, comparing server HTML with the rendered page
- Structured data findings, including anything marked up that the visible page does not support
- An entity consistency table listing every conflicting description we found, where it lives and who can change it
- A prioritised remediation backlog with effort marked and in-house items identified
- A one-page summary written for the person who will not read the rest of it
The obvious question
How this differs from a technical SEO audit
They overlap on technical findings and diverge on almost everything else. If your fundamentals have never been audited, the conventional one usually comes first.
| Dimension | Technical SEO audit | AI search audit |
|---|---|---|
| The question | Can search engines crawl, render, index and rank this site | Can AI systems reach this content, and what do they say when asked |
| Evidence base | Crawl data, log files, Search Console, Core Web Vitals | Crawler access testing plus a coded set of assistant responses with transcripts |
| User agents reviewed | Googlebot and Bingbot, with rendering behaviour | Search crawlers plus GPTBot, PerplexityBot, ClaudeBot, Google-Extended and user-triggered fetchers |
| Content review | Titles, headings, duplication, internal linking, index bloat | Whether passages survive extraction, whether claims are dated and sourced, whether the entity resolves cleanly |
| Off-site work | Link profile and referring domains | How consistently the business is described across sources you do not own |
| Output | A prioritised technical backlog | A prioritised backlog plus a measured baseline you can re-run |
| Honest limitation | It cannot tell you whether the content is any good | It cannot tell you why a model said what it said, or what it will say next month |
Scope
What the audit establishes, and what it cannot
Both lists are published before you buy, because the second one is the part that separates a diagnostic from a sales document.
What it tells you
- Exactly which AI crawlers can and cannot reach each part of your site, with the response codes as evidence
- How often your business is named across a defined prompt panel, and how often a competitor is named instead
- Whether you are cited with a source or only described, reported as two separate figures
- Which specific statements about your business are inaccurate, and which source we believe they came from
- Which pages fail an extraction test, and the specific edit that fixes each one
- Where your entity data disagrees with itself across the sources a machine can reach
- Whether an ongoing programme is justified for you, including when the answer is no
What it cannot tell you
- Why a model produced a particular answer, because these systems do not expose that
- Whether any specific future answer will name you, in any assistant, at any point
- A single blended AI visibility score, because averaging across models and prompts destroys the finding
- How much revenue an assistant sent you, where the referrer is stripped before the visit arrives
- Anything about what a model was trained on, which is not observable from outside the lab
- Whether a change will work before it has been made and re-measured
Questions
What buyers ask before commissioning one
How is this different from the SEO audit we had last year?
A technical SEO audit asks whether search engines can crawl, render, index and rank the site. This asks a different question: whether AI systems can reach the content, and what they actually say when asked about your business and your category.
The evidence base differs accordingly. An SEO audit runs on crawl data, log files and Search Console. This runs on crawler access testing across AI user agents, plus a coded set of assistant responses with the transcripts attached. There is overlap in the technical findings and almost none in the rest.
What stops this being generic advice with our logo on it?
Every finding has to be traceable to evidence in the appendix. A recommendation about crawler access points at a response code we recorded. A recommendation about how you are described points at the transcript where it was described that way, and at the source we believe it came from.
If a finding cannot be traced to something specific we observed on your site or in a response, it does not go in the report. That rule removes most of what fills a generic audit.
How long does it take and what do you need from us?
Four weeks is the normal shape, and most of that is running the prompt panel properly rather than waiting on anyone. We need read access to Google Search Console and Bing Webmaster Tools, a contact who can answer questions about the business, and confirmation of your competitor set.
We do not need site access to complete the audit. Access to the CDN or hosting configuration speeds up the crawler diagnosis considerably, but everything can be established externally if that is easier.
What happens at the end?
You get the report and the raw data, and you own both. Roughly half the fixes on a typical backlog are things an in-house team can do without an agency, and we mark which ones those are rather than bundling everything into a proposal.
If we think an ongoing programme is not justified, we say so in the report. That happens fairly often, usually where the technical and content fundamentals need attention first, and it is a more useful answer than a retainer.
Will the audit tell us why a model said something about us?
No, and no audit can. These systems do not expose why a particular answer was produced, and anyone reconstructing that is guessing.
What we can often do is trace an inaccurate claim back to a source that is still publishing it: an old directory entry, a stale profile, outdated coverage, or a page on your own site describing a service you stopped offering. That is a hypothesis with evidence behind it, and it is usually correctable.
Start with the diagnostic, not the retainer
Fixed scope, fixed price, four weeks, and you keep the panel and the raw data at the end. If the finding is that you should spend your budget elsewhere, that is what the report will say.
If we don't deliver the work we agreed to deliver for reasons within our control, you don't pay for the undelivered work. Read our guarantee
Related
Where to go next
- the retainer this audit feeds
- a conventional SEO auditDifferent question, different evidence base. Often worth doing first.
- ongoing visibility measurementWhat the frozen prompt panel is designed to support.
- AI Overviews and your click-through rateThe Search Console side of the same diagnosis.
- Bing and Copilot indexingIncluded as standard in the access review.
- how we run an engagement
Last updated · Reviewed by Zubair Afzal