How to Run an AI Visibility Audit in One Week
Most teams discover their AI visibility problem by accident — a prospect mentions that ChatGPT recommended a competitor. This is a structured way to find out on purpose, in five working days, without buying anything first.
TL;DR
- Day 1 — establish a baseline: do assistants name you at all?
- Day 2 — check technical access: can AI crawlers reach and parse your site?
- Day 3 — audit your content for quotability.
- Day 4 — analyse who gets cited instead of you, and why.
- Day 5 — prioritise fixes by effort against likely impact.
You need a spreadsheet, access to two or three assistants, and the ability to read your own server logs or CDN analytics.
Day 1 — Baseline measurement
Write thirty questions a buyer would ask, spanning informational, comparative, and purchase intent. Do not include your brand name in any of them.
Run each question through the assistants your buyers use, in clean sessions with memory disabled. Record for each response: were you mentioned, in what position, how were you framed, who else appeared, and which sources were cited.
By end of day you should be able to state a baseline plainly: "We appear in roughly N of 30 category questions, usually in third or fourth position, and Competitor X appears in most of them."
If that number is zero, the rest of the week is about diagnosing why — and the cause is usually on Day 2 or Day 4.
Day 2 — Technical accessibility
An assistant that cannot fetch your page cannot quote it. Check, in order:
robots.txt. Are AI crawlers permitted? Note that blocking them is a legitimate strategic choice — but make it a decision, not an accident. Many sites block bots they never meant to block, often via an inherited rule or a security plugin's default.
Server logs. Filter for AI crawler user-agents over the last thirty days. Are they arriving? Are they getting 200s, or 403s and 429s? A rate limiter or WAF quietly returning 403 to a crawler is one of the most common invisible causes of zero visibility.
Rendering. Fetch a key page with JavaScript disabled. If the content is not in the initial HTML, assume it may not be read. Not every system executes JavaScript, and those that do may not wait for slow client-side rendering.
Structured data. Validate your Organization and page-type schema. Confirm every sameAs URL resolves.
llms.txt. Optional, cheap, and increasingly consulted. Worth having.
Day 3 — Content quotability
Take the ten pages you most want cited and read each one as if you had to extract a single quotable paragraph.
Ask of each:
- Does it answer a real question in the first hundred words, or does it warm up for three paragraphs first?
- Are there self-contained passages that survive being lifted out of context? Anything starting "as we saw above" does not.
- Does it contain original evidence — your own data, benchmarks, methodology — or does it restate what ten other pages already say?
- Is authorship attributed to a named person with real credentials?
- Is it current, and does it say so?
Content that only paraphrases the consensus gives an assistant no reason to prefer you. Original evidence is the single strongest differentiator, because it exists nowhere else.
Day 4 — Competitive analysis
For every question where a competitor appeared and you did not, work out what they have that you lack. Open the sources the assistant cited and compare.
The answer is usually one of:
| Pattern | What it looks like |
|---|---|
| Third-party presence | They appear on review sites, listicles, and comparison pages; you do not |
| Original research | They published data everyone else now cites |
| Structural fit | Their page directly answers the question; yours buries the answer |
| Entity clarity | Their brand is unambiguous; yours collides with another organisation |
| Community footprint | They are discussed in forums and communities the models draw on |
Third-party presence is the one that surprises people most. Assistants frequently ground category recommendations in independent sources rather than vendor sites, which means the fix lives partly off your own domain.
Day 5 — Prioritise
Score each finding on effort and expected impact, then sequence.
Usually fast and high value: unblocking crawlers, fixing broken or missing Organization schema, correcting dead sameAs links, restructuring existing high-intent pages so the answer comes first.
Slower but higher ceiling: publishing original research, earning presence on the third-party sources that actually get cited, building a named-author byline with real credentials.
Deprioritise: chasing exotic schema types, minor page-speed gains, and volume content with nothing new in it.
Set a re-measurement date four weeks out, and re-run Day 1 exactly as before. Same questions, same method. Without that, you will never know which fix worked.
What the audit will not tell you
One audit gives you a snapshot. It cannot distinguish a genuine improvement from ordinary run-to-run variation, and it cannot tell you whether a competitor's gain is a trend or a blip. Both require repeated measurement over time.
Run the audit to find out whether you have a problem and where it lives. Move to continuous measurement once you start acting on it.
Frequently asked questions
Can I do this without a tool? Yes, for a first pass. Manual auditing does not scale to continuous monitoring, but it is a perfectly good way to size the problem.
How many questions is enough? Thirty is a reasonable minimum. Below twenty, individual oddities distort the picture.
Should I block AI crawlers? Only deliberately. Blocking protects content from training use but also removes you from answers. Most brands that want visibility should allow them on public marketing content.
How soon can I expect change? Technical fixes can show within weeks. Authority and third-party presence take a quarter or more.
Published by the SIQA Editorial Team.