How do we audit our existing content for AI assistants?

The website has been gathering pages for years: service pages, blog posts, news items, a few guides someone commissioned in a good year. Nobody can say which of them do any work. Then a prospect mentions that they asked ChatGPT for firms like yours and the answer named two competitors, and the first instinct is to write something new.

Resist it for a month. Ranking on Google is no longer enough on its own: Ahrefs found in March 2026 that only 38% of the pages cited in Google's AI Overviews ranked in the top ten for the search, down from 76% in July 2025, because Google now runs several related searches for one question. Most of what an assistant needs is probably already somewhere on your site, but out of date, badly arranged, unreadable to AI crawlers or buried under pages that say nothing.

Before you write, find out what you have, which pages AI assistants already use, which are worth fixing and which should go. That audit needs a spreadsheet, a scoring method and, in our experience, about a day for every twenty pages. Below are our method, a template with a scored example, and a six-month plan for a firm without a content team.

#Why should we audit before writing anything new?

Because the pages you already have are usually closer to working than a blank page, and weak pages on a subject blur which page is yours.

Content written to rank on Google tends to fail in the same ways when an assistant reads it: it says nothing specific, shows no sign of who wrote it, repeats what competitors say and is out of date. Google's guide to its AI search features, last updated in July 2026, says content with first-hand expertise that could not come from anyone else counts for most, and warns that creating separate pages for every variation of a question can break its scaled content abuse policy. So an audit that merges and deepens usually does more than one that adds. Why ChatGPT never quotes our blog explains the five failures in detail.

Scale is no reason to skip it. For TCC Group, a regulatory compliance group with three brands, the audit meant "624 pages reviewed, each given a decision: migrate, merge or retire." What it found was not weak writing: "What failed was underneath: a site organised around internal teams rather than the questions buyers ask, pages that half-answered the same question so neither was cited, and machine-readable data that declared a name, a logo and nothing else."

#What does the scoring method look like?

Score each page from one to five on the same five reasons we use in why ChatGPT never quotes our blog, where one means no problem and five means the problem is severe. Add the five scores for a total between 5 and 25.

Then read the total against our working bands:

These bands are our judgement from audits we have run, not a published standard. Where two pages tie, fix point of view and specificity first, because they move the result most, then evidence of expertise, then structure, then authorship.

Two more columns sit alongside the score. Currency records when the page was last reviewed; see how old is too old for AI assistants for review cycles. AI use records whether assistants already draw on the page: a poor page that AI already cites is fixed carefully and in place, and a good page that nothing uses may have an access or linking problem.

#How to run the audit, step by step

#Step 1: List every page with its subject, traffic and importance

Export the site's pages from your content system or from Google Search Console. Add columns for the subject in plain words, the last twelve months' traffic and strategic importance, scored high, medium or low by asking three questions: do buyers ask about this, is it central to what you sell, and do we know something about it that competitors do not? This list is the audit's spine.

#Step 2: Check that GPTBot and other AI crawlers can reach your pages

An assistant cannot use a page its crawler cannot reach or read. Open yoursite.com/robots.txt and look for rules naming these crawlers:

OpenAI says a site must allow OAI-SearchBot to appear in ChatGPT search. Then check your CDN: Cloudflare changed its default to block AI crawlers in July 2025, so a site behind Cloudflare may be blocking them without anyone having decided to. Finally, view the source of three key pages (not the rendered page) and search for the main heading and first paragraph. Vercel and MERJ found that GPTBot, ClaudeBot and PerplexityBot fetch JavaScript files but do not run them, so text that only appears after scripts run is invisible to them. Our technical checklist for an AI-ready website covers the rest.

#Step 3: Find out which pages AI assistants already use

Use three sources, because none is complete on its own.

• Prompt checks. Keep a fixed list of 20 to 50 buyer questions and ask each several times a month in ChatGPT, Gemini, Claude, Copilot and Perplexity, in a clean session. Record which of your pages is cited. Answers vary from run to run, so count a page as used only when it appears repeatedly.
• Analytics. Since 13 May 2026, GA4 has a default "AI Assistant" channel covering ChatGPT, Gemini, DeepSeek, Copilot and Grok. Claude and Perplexity are not included, so add a report filtered on referrers such as perplexity.ai and claude.ai. Clicks from AI Overviews and AI Mode are counted as Organic Search; Search Console's generative AI performance reports, available to all sites since 31 August 2026, show those impressions by page.
• Server logs. Ask whoever hosts the site for a month of logs and count hits by user agent: GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-User, PerplexityBot and Perplexity-User. A ChatGPT-User or Claude-User hit means a person's assistant fetched the page to answer a question.

Put the result in the AI use column as a short note, for example "cited 3 of 4 runs; 12 referrals; crawled by OAI-SearchBot".

#Step 4: Score the five reasons

Read each page against the rubric above and score it one to five on each reason. An AI assistant can draft these scores if you give it the page, the rubric and the top competitor pages, which saves time. A person who knows the subject and the competitors must check every score, especially point of view, which is the hardest to judge and the most valuable.

#Step 5: Add the currency check

Record the last reviewed date and flag any page tied to revenue that is more than three months past review, and any other page not reviewed in a year. Judge the content, not only the date: a three-year-old explanation of a principle that has not changed can be fine, while a one-year-old page quoting a superseded regulation is not.

#Step 6: Decide what happens to each page

Combine the total with strategic importance, using the decision table below. Write one decision per page: maintain, fix, rewrite, merge or retire.

#Is there a free content audit template for AI search?

Yes: copy these columns into a spreadsheet. The row below is a fully scored example for an imagined firm of chartered surveyors, with obviously illustrative figures.

#How do we turn scores into decisions?

Put the total on one axis and strategic importance on the other. The time estimates are from our own work and will vary with the subject.

A good page on a subject nobody buys from you is still not worth maintaining. Merge where two pages answer the same question, and redirect the old address to the one you keep. A site with fewer, stronger pages reads as more expert than one with many weak ones, for reasons set out in why ChatGPT ignores our website when we have hundreds of pages.

The audit will also show gaps: important subjects with no page at all. Those go on the new content list, and they should be the only things on it until the rebuilds are done.

#What does a realistic six-month plan look like?

Assume around a hundred pages and no content team. Each month has one job, so nothing overlaps.

How to organise our content so AI assistants see us as the expert explains how the rebuilt pages should link together, and how to find out whether ChatGPT is recommending our firm turns the baseline into a monthly routine.

#What to do this month

1. Export the list of pages and add subject, traffic and strategic importance for each.
2. Check robots.txt, your CDN's bot settings and the page source of three key pages, and fix anything that blocks AI crawlers.
3. Record the AI use baseline: run your fixed list of buyer questions, set up the AI referral report in GA4 and ask your host for a month of server logs.
4. Score the twenty most important pages on the five reasons and the currency check, using the template above.
5. Do the quick wins on every high-importance page: dates, figures, a direct opening answer and a named author.
6. Book practitioner time now for the first targeted fixes next month.

#Frequently asked questions

#How long does a content audit take for a 200-page B2B website?

In our experience, about twelve working days for one person: a day to export the pages and check access, a day to set up the AI use checks, and about a day for every twenty pages scored. Start with the twenty that matter most, which takes a day or so and shows the patterns that make the rest faster.

#Should I delete or noindex weak pages before optimising for AI search?

Decide page by page, and do not do it blind. Check AI use and inbound links first: some weak-looking pages are still cited or linked. Merge overlapping pages and redirect the old address; delete pages that are wrong and have no links or traffic; use noindex only for pages visitors need but searchers do not, since it keeps a page out of Google's index without removing it from your site. On arum's rebuild, "Every old address redirected, so years of rankings carried over."

#Can a tool or an AI assistant do the scoring for us?

Partly. Tools measure structure, length, dates and crawler access reliably. An AI assistant can draft the five scores if you give it the page, the rubric and the competitor pages, and it is often a good first pass. A person who knows the subject and the competitors must check every score, because judging point of view needs knowledge of what others have published.

#Do we need to score every page, or can we start with some?

Start with the twenty pages that matter most: your core services and the pages that still bring in traffic or enquiries. Scoring those shows the site's patterns and often changes the priorities. A full audit can follow over a quarter, but there is nothing to gain by waiting for it before fixing the obvious pages.

#What if we have no data of our own for the point of view score?

Point of view can come from judgement as well as data. A page that explains why the standard approach fails in a common situation, or compares options nobody has compared, is distinctive without a single figure. Ask your practitioners what they believe that competitors would argue with.

#How often should we repeat the audit?

A full pass once a year, and a quarterly check of the twenty most important pages for currency and AI use. Score every new page against the same five reasons before it is published, so the standard holds.

If you would rather have the audit run for you, the Paid Discovery takes one week and ranks every finding, with fixes written so your own team can ship them; for TCC Group it produced "1 week to a plan TCC's own team built the new site from." Or start with a free Growth Review: 30 minutes with David Adams and three fixes in writing.

#Sources