The half-day version
For a B2B firm with a working site, decent rankings, and no idea whether ChatGPT has ever named them.
- 1Run ten real buyer questions through ChatGPT, Perplexity and Google. Log every firm cited. That is your baseline, and it takes twenty minutes.
- 2Check you are retrievable: view source on your three money pages and confirm the words are in the HTML, then read robots.txt for accidental AI crawler blocks.
- 3Score those three pages against the 20-point scorecard below. Anything under 12 is why you are not being quoted.
- 4Find the evidence gap: list every number on your site that a competitor could not have written. If the list is empty, that is the whole problem.
- 5Search your own company name and see what the web says about you outside your own domain.
- 6Write the fixes down in impact order and do the top three this month, not all twenty.
The short answer
An AI search audit asks five questions of your site, in this order: are you being named today, can a machine read your pages at all, can it lift a clean answer out of them, is there anything on the page it could only get from you, and does the rest of the web associate your name with your specialism. Fail the first two and nothing else matters. Most B2B sites pass the first two easily and fail the last three completely.
It is not a rebrand of a technical SEO audit. A crawl tool will tell you your title tags are duplicated. It will not tell you that every paragraph on your services page could have been written by any of your competitors, which is the actual reason a model paraphrases you instead of citing you.
What you are actually auditing
A traditional audit measures a page's ability to hold a position. This one measures a paragraph's ability to survive being ripped out of context. When an assistant answers a question, it does not read your page the way a prospect does. It pulls chunks, drafts a reply, and attaches links to whichever chunks carried the load. Your unit of optimisation shrank from the page to the passage.
So the audit is not really about your website. It is about whether the individual passages on your website can stand up alone, say something specific, and require attribution when repeated. That is a writing and evidence problem far more often than it is a technical one.
"A crawl tool grades your plumbing. This audit grades whether you have anything worth quoting once the plumbing works."
Baseline: are you named at all?
Do this before you change anything, because without a before you will spend a quarter improving pages and have no idea whether it moved. Twenty minutes, a spreadsheet, ten questions taken straight from your last five sales calls.
- Use the buyer's words, not yours. Nobody types "B2B demand generation solutions". They type "why are we getting clicks from LinkedIn but nobody books a call".
- Run each question three times. Answers vary between runs. One absence proves nothing; three in a row is a signal.
- Log who is cited, not just whether you were. The named sources tell you what a winning page looks like for that question, and it is usually not who you assumed.
- Ask the assistant about you directly. "What does {your firm} specialise in?" A wrong or vague answer means your entity is muddled, which is a different fix to a content fix.
Keep it in a sheet. Re-run the identical set monthly and never change the wording, or the comparison is worthless.
Question:
Engine (ChatGPT / Perplexity / AI Overview):
Date run:
Sources cited (in order):
Were we named? Y / N
Which page of ours was quoted:
Was the answer about us accurate? Y / N
Notes: what the top-cited page does that ours does not
The last line is the whole audit in one field. Nine times in ten the answer is: it states a number in the first paragraph and we state a benefit in the first paragraph.
Retrievability: can a machine read the page?
This is the short, boring section that occasionally explains everything. Work through it once and you probably never need to again.
- View source on your three most important pages. If the body copy is not in the raw HTML and only appears after JavaScript runs, some crawlers get an empty shell. Prerendering or server rendering fixes it.
- Read robots.txt line by line. GPTBot, PerplexityBot, ClaudeBot, Google-Extended, CCBot. Blocking them is a legitimate business decision. Blocking them because a plugin defaulted to it is not.
- Confirm the pages are indexed. Search Console coverage, or a plain site: search. Assistants build answers from an underlying index. Not in the index, not in the answer.
- Check every page has a unique title and description. Duplicated metadata makes pages look interchangeable, and interchangeable pages get skipped.
- Find the answers hiding behind forms. A gated PDF cannot be quoted. Publish the substance on the page and gate the template or the tool instead.
- Check your numbers are text. A benchmark table exported as a PNG is invisible. That one catches almost everybody.
Would rather focus on leads than run this yourself?
Book a 30 minute call and tell us what your pipeline looks like. We will tell you straight if we can help fill it.
Quotability: score your pages properly
Take your three money pages and read them as a machine would: one section at a time, out of order, with no memory of what came before. Then ask the only question that matters. Could this passage be quoted on its own and still be saying something?
"We take a data-driven approach to B2B advertising, tailoring every campaign to your unique objectives and delivering measurable results across the channels that matter most to your business."
Four clauses, no claim, nothing to attribute. Any competitor could publish it unchanged.
"On UK LinkedIn campaigns we run for niche B2B firms, a landing page click typically costs a few pounds and CPC climbs sharply once you narrow to senior job titles. Judge the account on cost per qualified pipeline, not cost per click."
Scoped, specific, and it takes a position. That needs a source, and the source is you.
Mark up a printout if it helps. Highlighter through every sentence that survives being lifted. If a page comes back with three highlights out of forty paragraphs, you do not have a technical problem, you have a writing problem, and no amount of schema will paper over it.
Evidence: is there anything only you have?
Models synthesise. Ten pages saying the same thing in different words give a model no reason to name any of them. One page carrying a figure nobody else has forces attribution, because the claim needs a source. So audit your site for evidence the same way you would audit it for broken links.
- Count the first-party numbers. Figures from your own accounts, your own clients, your own tests. Be strict: a statistic you quoted from someone else's blog does not count, they get the citation, not you.
- Check every worked example has real figures. Spend, cost, what you changed, what happened next. "We significantly improved lead quality" is unquotable and everyone knows it.
- Look for a named method. A framework with a name and a one-line definition is easy to attribute. An unnamed process gets described generically and credited to nobody.
- Find the hedging. Every "it depends" that never resolves into an answer is a passage that will not be cited. Say what you think, then qualify it.
- Verify the sample and the period on any data you publish. Numbers without a stated basis get discounted by careful sources, and increasingly by models too.
The guardrail here is non-negotiable: only publish figures you could defend line by line if a client challenged them. Invented statistics get contradicted by other sources, and a model that finds a conflict usually cites neither party.
Off-site: what the web says about you
Half of an answer engine's opinion of you is assembled from sources you do not own. This part of the audit takes an hour and is the part most people skip.
- Search the roundups. "Best {your service} agencies UK" and the sector variants. Those pages are exactly what gets read when someone asks for a recommendation. Note every list you are missing from.
- Check the communities. Reddit and niche forums get retrieved heavily. Is your firm mentioned anywhere by anyone who is not you?
- Audit your entity consistency. Same company name, same one-line description, same founder name on your site, LinkedIn, directories and Companies House. Inconsistency dilutes the association between you and your specialism.
- Inventory transcribed material. Podcasts, webinars and YouTube videos with real descriptions and transcripts are large bodies of attributed first-person text. Most B2B firms have recordings sitting unused.
The 20-point scorecard
One point each, scored per page for the content rows and per site for the technical ones. Run it on your three most commercially important pages. Under 12 and you can safely assume you are being paraphrased rather than cited.
| Area | The check | Pass looks like |
|---|---|---|
| Retrieval | Body copy present in raw HTML | Words visible with JavaScript disabled |
| Retrieval | AI crawlers not blocked by accident | robots.txt reviewed and deliberate |
| Retrieval | Page indexed in conventional search | Appears in Search Console coverage |
| Retrieval | Unique title and meta description | No duplicates anywhere on the site |
| Retrieval | Answer is not gated | Substance on the page, tool behind the form |
| Structure | Direct answer in the first 100 words | A reader gets the answer before the story |
| Structure | Headings phrased as questions | Headings match how buyers ask |
| Structure | Sections stand alone | No 'as we covered above' |
| Structure | Comparisons in a real table | Marked-up table, not an image |
| Structure | Numbers written as text | Every figure appears in a sentence |
| Structure | Genuine FAQ block | Real buyer questions, answered plainly |
| Structure | Published and reviewed dates | Visible on the page |
| Evidence | At least one first-party figure | From your accounts, not a third party |
| Evidence | A worked example with real numbers | Spend, result, and what changed |
| Evidence | A named method or framework | Named and defined in one line |
| Evidence | A clear stated position | You say what you think |
| Evidence | Sample size and period stated | Any data has a basis attached |
| Off-site | Named in at least one roundup | A list you did not write |
| Off-site | Consistent entity details | Same name and description everywhere |
| Off-site | Transcribed material published | Podcast, video or webinar text |
The scoring is deliberately blunt. The value is not the number, it is that the failed rows are already a prioritised task list, and the Evidence block is where almost every B2B firm loses its points.
Fixing it: the first 30 days
An audit that produces a 40-item backlog gets ignored. This is the order I would work in with one afternoon a week.
- Week 1Baseline and unblock
Run the ten-question panel across all three engines and log it. Then clear the retrieval rows: robots.txt, rendering, indexing, duplicate metadata. These are cheap and they gate everything else.
Output: A dated scoreboard and a site that can actually be read. - Week 2Rewrite the top of your three money pages
Move the direct answer into the first hundred words, turn headings into the questions buyers ask, break the 'as mentioned above' dependencies and pull every number out of images into text.
Output: Three pages that can be lifted from without interpretation. - Week 3Close the evidence gap on one page
Pick the page closest to a buying decision and put something in it nobody else has: your own anonymised benchmark with sample and period, or a worked example with real spend and outcome. Name the method while you are in there.
Output: One page that forces attribution instead of paraphrase. - Week 4Repair the off-site picture
Fix inconsistent entity details across your site, LinkedIn and directories. Pitch yourself to the two roundups that already rank for your category, and publish one transcript you already own.
Output: Consistent identity and mentions in sources models already read.
Then re-run the panel and resist reading too much into it. One month rarely moves a citation. What it does tell you is whether the rewritten pages are being retrieved at all, and that is the thing worth knowing before you commit a quarter.
What the audit should ignore
- An "AI readability score" from a tool. Nobody has a validated model of how citations are chosen. A composite number invented by a vendor is not evidence, and it will send you optimising for the vendor's heuristic.
- Auditing every page. Three to five pages carry your commercial weight. Grade those. The rest of the site is not what is holding you back.
- llms.txt. A proposed standard the major assistants have not committed to. Add it in ten minutes if you want, but do not put it on the list of fixes that matter.
- Word count targets. Length is not the variable. A tight 900-word page with one real number outperforms 3,000 words of context every time.
- Citation-monitoring subscriptions on day one. Run the manual panel for a quarter first. If it turns into a real channel, then pay for tooling.
"Most AI search audits fail at the same row: there is nothing on the site a model could only have got from you."
Want more qualified leads without adding headcount?
Book a 30 minute call. Tell us where your enquiries come from today and we will give you an honest read on what to do next.
Read next
Common questions
If you want ads and content working together to bring in leads
- LinkedIn, Meta and Google campaigns run against a named target account list.
- Creative built from your own material, shipped and tested every month.
- Tracking and attribution wired up so you can see what is actually working.
- Content and landing pages written to answer the questions your buyers ask.
- Reporting on cost per qualified pipeline rather than cost per click.
Starter £1,250/mo, Growth £1,725/mo, Scale £4,025+/mo. Content work sits inside the retainer where it is in scope. Anything outside that is quoted properly before it starts.