How to Audit Your Website for AI Content That Needs Fixing

Crawl or export every page, mark which ones AI drafted or rewrote, then score each on five checks: facts still true, claims supported, not duplicated across pages, genuinely useful, and free of AI leftovers. Fix, merge, redirect or remove in order of traffic and risk, starting with pages that make claims or quote prices.

The pages that need fixing first are rarely the obviously robotic ones. A clumsy paragraph costs a little trust; a confident, well-written page promising a turnaround you stopped offering costs you a client. So the audit starts with what each page says, and only then with how it reads.

Follow me on Instagram@sagnikteaches

Build the page inventory in one spreadsheet

A content audit is only as good as its list of pages. The example here is an illustrative eight-person translation agency whose site grew from 20 pages to 68 over two years, with most of the growth drafted by AI. Its inventory came from three places:

Connect on LinkedInSagnik Bhattacharya

• A crawl. Screaming Frog's SEO Spider is free for up to 500 URLs, which covers most small business sites. It lists every page it can reach with its title, word count and status code. A paid licence removes the limit and adds features such as custom searches of page source and a Search Console connection.
• Search Console. The Performance report gives clicks and impressions per page for the past 12 months. Its tables and exports stop at 1,000 rows, which is plenty for a 68-page site. The Page indexing report shows which pages Google crawled and chose not to index, a useful hint about pages it considers low value.
• Your own records. The content calendar, CMS revision history and whoever wrote what. This is how you fill the column that matters most: how much AI was involved.

Six rows from the agency's sheet, filled in:

Subscribe on YouTube@codingliquids

Where nobody remembers whether AI was involved, mark it "unknown" and audit it anyway. The scoring below doesn't depend on who wrote the page, only on whether it's right and worth keeping. If you'd rather start with a broader SEO health check than a content audit, auditing your website's SEO with AI in one afternoon covers indexing, speed and enquiry routes.

Leftovers that show a draft was never finished

Before scoring, spend half an hour hunting for text that plainly shouldn't be on a live page. Search your CMS, or use a site search in Google (the phrase in quotes after site:yourdomain.com), for strings like these:

• "As an AI" and "As of my last update"
• "Certainly! Here's" and "I hope this helps"
• Square-bracket placeholders: "[insert", "[Company Name]", "[X]"
• "Lorem ipsum"
• Stock openings your team never writes, such as "In today's globalised world"
• Any FAQ question that appears word for word on more than three pages
• Image alt text that describes the prompt rather than the picture, such as "professional translator at desk, photorealistic, 4k"

The agency's search turned up three service pages still containing "[insert turnaround time]", one article ending "I hope this helps! Let me know if you'd like any changes", nine articles opening with the same globalised-world sentence, and a 14-page run of language pages sharing an identical FAQ block. None of these had been noticed in two years, because nobody reads their own site the way a stranger does. Fix these on the day you find them; they don't need scoring.

Score every page on five checks

Each page gets 0, 1 or 2 on five checks, for a score out of 10. The checks are about substance, not style:

1. Accurate now. Prices, turnaround times, services, tool names and staff details all still true. Stale facts are the most common failure on AI-drafted pages, because nobody revisits them; catching outdated information in AI answers covers the categories that date fastest.
2. Claims supported. Every "100%", "fastest", "500 linguists" and "accepted everywhere" has evidence you could show. The method is in backing up the claims AI writes.
3. Unique. Not a near-copy of another page on your site or someone else's. A plagiarism checker with a private index finds both; see checking AI content for plagiarism.
4. Useful. Answers a real question with specifics a reader couldn't get from any other agency's page: your process, your numbers, your examples.
5. Clean. No leftovers, no filler, and it sounds like you.

The agency's scores for the six pages above:

Page                               Acc  Sup  Uni  Use  Cln  Total  Notes
/services/certified-translation     0    0    2    1    2    5     "within 24 hours" (it's 2-3 days);
                                                                   "accepted by all authorities"
/languages/english-spanish          1    1    0    1    1    4     same text as 23 other language pages
/languages/english-czech            1    1    0    0    1    3     same, and almost no traffic
/insights/what-is-transcreation     2    2    1    1    0    6     generic opening, no examples
/insights/translation-memory-tools  0    2    1    1    1    5     calls Phrase TMS "Memsource"
/about                              2    2    2    2    2   10     keep

Score the page titles and meta descriptions too, not just the body. Google's guidance on generative AI content says automatically generated titles, descriptions and structured data should meet the same standards as the page itself, and AI-written meta descriptions are where promises tend to hide, because nobody reads them after publishing. The Czech page's description read "Same-day English to Czech translation by certified experts. Guaranteed accuracy." Neither claim was true, and it was the first thing anyone searching saw.

The translation-memory article shows why accuracy is checked separately from how a page reads. It's fluent and fairly useful, but it describes Memsource, a product that has been called Phrase TMS since the company unified its brand in 2022. To a translator reading the agency's site, that one name says "nobody has looked at this in years".

Letting AI do the first pass of scoring

Scoring 41 pages by hand is the slowest part of the audit, and AI can take a first pass at three of the five checks: unique, useful and clean. It can also list every claim on a page for a person to verify. What it can't judge is accuracy, because it doesn't know your current prices or turnaround times, so supply those or leave that column to a person. A prompt that works page by page:

Below are our current facts, then the text of one web page.
1. List every factual claim on the page (numbers, times, prices,
   guarantees, tool names, staff numbers). For each, say whether our
   facts support it, contradict it, or don't mention it.
2. Score the page 0-2 on UNIQUE (does it read like a template that
   could apply to any language or any agency?), USEFUL (specific
   answers a customer couldn't get elsewhere?) and CLEAN (filler,
   leftovers, generic openings?). One sentence of reason each.
Don't rewrite anything.
[facts sheet]   [page text]

An illustrative result for the English-to-Spanish page:

Claims: "24-hour turnaround" - contradicted (facts: 2-3 working days)
        "certified translators" - supported
        "over 500 linguists" - contradicted (facts: about 60)
UNIQUE 0: the page reads as a template; only the language name is
          specific to Spanish.
USEFUL 1: explains certification, but no Spanish-specific detail.
CLEAN  1: opens with a generic sentence about global markets.

That's close to what the agency's own reviewer concluded, in a fraction of the time. The reviewer still reads each result and signs the score off, because a model can miss a claim or call a genuinely specific paragraph generic. Used this way, the first pass cut scoring from about 6 minutes a page to about 3.

Keep, fix, merge, redirect or remove

The score tells you how bad a page is; traffic and links tell you how much it matters. Combine them:

When you merge, use a permanent (301) redirect from the old address to the page that replaces it, so visitors and any links still arrive somewhere useful. When you remove a page nobody visits and nothing links to, a plain removal is fine. Don't keep weak pages simply because they exist; a site with 49 strong pages is easier to keep accurate than one with 68 mixed ones.

This isn't about hiding AI from Google. Google's guidance says AI-generated content is fine when it's helpful and accurate, and that using generative AI to produce many pages without adding value for users may breach its spam policy on scaled content abuse. The language pages below are a textbook case of the second. For the fuller picture of what Google does and doesn't penalise, see whether Google penalises AI-written content.

What to do with 24 near-identical language pages

The agency's biggest problem was a set of 24 language-pair pages generated from one template in early 2025: "English to Spanish translation", "English to Czech translation" and so on, each around 500 words, differing only in the language name. Search Console told a clear story. Six pages brought in 91% of the set's clicks. Eleven had fewer than ten clicks in a year. Seven sat in the "Crawled, currently not indexed" group in the Page indexing report.

The decision:

• Keep and rewrite the six pages with real demand. Each now carries specifics only that pair can have: which document types the agency's translators for that language usually handle, typical turnaround for that pair, how many translators work in it, and a short worked example of a common job.
• Merge the other 18 into one "Languages we translate" page, with a table listing each pair, typical turnaround and specialist areas.
• Redirect all 18 old addresses to the new page.

Rewriting six pages properly took longer than generating 24 badly, which is the point. A reader looking for Spanish translation now finds a page that answers their question, and the agency has 18 fewer pages that could go stale.

Fixing a page without starting again

Most pages with a score of 4 to 7 don't need a rewrite from scratch. They need their facts corrected, their claims narrowed and their filler replaced with specifics. The certified translation page, before and after:

Before: We provide 100% accurate certified translations accepted by all authorities worldwide, delivered within 24 hours. Our team of over 500 native linguists ensures unmatched quality.

After: Our certified translations come with a signed statement from the translator confirming the translation is complete and accurate. Most organisations accept this, but requirements differ, so check with whoever will receive the document; we're happy to confirm before you order. Standard turnaround is two to three working days for up to 2,000 words, with a 24-hour option on request. We work with around 60 qualified translators, each translating into their native language.

Five problems fixed in one paragraph: "100% accurate" (unsupported), "all authorities worldwide" (false and risky), "within 24 hours" (not the standard service), "500 linguists" (the real figure is about 60) and "unmatched" (filler). The rewrite is longer, more specific and more believable. When AI helps with fixes like this, give it the facts first:

Rewrite the paragraph below using ONLY these facts:
- certified translations include a signed accuracy statement
- acceptance varies; we confirm requirements before ordering
- standard turnaround: 2-3 working days up to 2,000 words;
  24-hour option on request
- about 60 qualified translators, native-language only
Remove any claim not supported by these facts. Keep it under 90 words.
[paste the current paragraph]

Once a fixed page goes live, route future edits through the same checks as new content; the draft, check and sign-off workflow shows how.

Deciding the order when there are 40 pages to fix

Fix in order of risk multiplied by exposure. Pages that quote prices, turnaround times, guarantees or legal points come first, because they create obligations and complaints. Within that group, the busiest go first. The agency's order:

1. Service pages with claims or prices and real traffic (9 pages).
2. The six language pages being rewritten, and the new merged page with its redirects.
3. Pages with claims but little traffic. Low exposure, but a single customer relying on a wrong statement still matters.
4. Busy articles with stale tool names or generic content.
5. Everything else, merged or removed.

A page that's wrong and busy is worth an hour today. A page that's merely bland and unvisited can wait for the merge.

The agency's audit in hours and pages

Here's how the illustrative audit added up, start to finish:

• Inventory: crawl, Search Console export and AI-involvement column, 2 hours.
• Leftover search and same-day fixes: 1 hour.
• Scoring 41 AI-involved pages with an AI first pass, at about 3 minutes each: roughly 2 hours.
• Decisions and fix order: 1 hour.
• Fixing 9 service pages at about 45 minutes each: 6 hours 45 minutes.
• Rewriting 6 language pages at about an hour each, plus 2 hours for the merged page and 1 hour for redirects: 9 hours.
• Fixing or merging 12 articles at about 30 minutes each: 6 hours.

That's about 27 hours 45 minutes, spread over a month. At an internal cost of $60 an hour, roughly $1,670. The site went from 68 pages to 49: 18 language pages merged into one, and three thin articles combined into one stronger piece. Every remaining page now has a named owner and a "facts last checked" date in the spreadsheet.

Checking the fixes landed

An audit isn't finished when the edits are saved. Within a day of the changes going live, and again a few weeks later, run these checks:

• Every redirect works. Paste each old address into a browser and confirm it lands on the intended page, not the homepage or an error. A second crawl with the same tool lists any broken links the merges created inside your own site.
• Key pages are re-read by Google. In Search Console, inspect the rewritten service and language pages and request indexing, so the corrected turnaround times replace the old ones in search results sooner.
• Traffic moves where you expected. Four to eight weeks on, compare clicks for the merged language page with the 18 pages it replaced. A merged page should hold most of that traffic; if a particular pair's searches have dropped away, that pair may deserve its own page back.
• The facts sheet and the pages agree. Pick five pages at random and check their figures against the source of truth. If one disagrees already, the process has a gap.

Keeping the next audit small

The expensive part of this audit was two years of pages nobody had looked at since publishing. Four habits keep the next one to an afternoon:

• Record AI involvement when a page is published, in a CMS field or the content calendar, with the date and the editor's name.
• Keep one source of truth for prices, turnaround times and service details, and change it first when anything changes. Then search the site for the old figure.
• Check the 20 busiest pages every quarter. About 30 minutes, using the same five checks.
• Keep a glossary of current product and tool names, so renamed tools don't creep back in through new AI drafts.

The aim is a site where every page could survive a sceptical customer reading it closely, because some of them do.

Content audit questions

Should we label pages that AI helped write?

There's no general rule requiring labels on text that people have edited and checked. Google's own guidance suggests that explaining how content was made can help readers trust it, so a short note on how you produce content is reasonable. Images, video and audio are different: platform and legal labelling rules can apply there, so check them separately.

Will removing pages hurt our search traffic?

Removing pages that get no clicks and have no links pointing at them rarely costs anything. Pages with some traffic or links should be merged into the closest strong page and redirected, so visitors and links arrive somewhere useful. Watch Search Console for a few weeks afterwards and restore anything that turns out to have mattered.

Can AI help run the audit itself?

Yes, for the mechanical parts: sorting the inventory, spotting near-duplicate pages, finding leftover phrases and drafting fixes from facts you supply. It can't tell whether your prices, turnaround times or claims are still true, because it doesn't know your business. Keep that judgement with whoever owns the facts.

How often should the audit be repeated?

A full audit once a year is enough for most small sites, with a quick check of the 20 most-visited pages every quarter. Run an extra audit after any burst of AI-assisted publishing, or when prices, services or tools change, because that's when a lot of pages go stale at once.

Further reads

• How to Write Website Articles With AI That Still Rank on Google — Write replacement articles with AI that still rank.
• How to Write Your Website Copy With AI: Home, About, and Services — Rewrite core pages from your own facts, not generic filler.
• How to Build a Product Glossary AI Must Use in Every Draft — Stop renamed tools and services reappearing in new pages.
• What to Do When ChatGPT Gets Facts About Your Business Wrong — Fix what AI search answers say about your business too.
• How Small Translation Agencies Build an AI-Assisted Workflow — How a translation agency builds AI into its day-to-day work.
• How to Run a Monthly AI Quality Review in 30 Minutes — Keep a monthly check going so the next audit is smaller.
• How to Check AI Content for Plain English and Readability — A four-stage routine for checking AI drafts for plain English: readability scores, a jargon audit prompt, targeted rewrites and a five-minute reader test.
• AI Alt Text: How to Generate and Check Image Descriptions — Generate alt text with AI for product photos, banners and charts, then check every description with a five-point test before it goes live.
• AI Website Builder or Web Designer: What Each Really Costs — Builder prices for Wix, Framer, Durable and Hostinger, what designers charge according to Clutch, and a three-year cost that counts your own hours.
• How to Choose an SEO Agency That Understands AI Search — Gather your own AI search baseline, ask agencies the questions that expose GEO hype, score two proposals and know what honest monthly reporting looks like.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Google Search Central (guidance on using generative AI content, scaled content abuse); Search Console Help (export and table row limits, Page indexing report); Screaming Frog SEO Spider pages (free version crawl limit and licence features); Phrase (Memsource renamed Phrase).