GEO/AEO audit: why Google and AI disagree on your site

Last week I showed you the split: Google cut my reach by about 56 percent over the same weeks my AI citations climbed past 380. I ended that issue saying I would stop guessing and actually look. This week I did.

I ran a full audit of how every search engine, the classic ones and the AI ones, sees my site. I fed it real data from four different dashboards, checked it against what Google now says about AI search, and had Claude Code dig through my own code to find the causes behind what the dashboards were showing.

This issue is the walkthrough I am going to present to you below: what I fed it, how I ran it, and what it found. This week I also implemented the first fixes and pushed the code, and I am already seeing some slight early movement. I am leaving the measurement running all week and will report the real numbers next week. If you run a site, a newsletter, or a product, you can run this same process yourself, and I will show you how in the following paragraphs.

An SEO audit and a GEO audit ask different questions

A normal SEO audit asks one question: will Google rank this? A GEO ( generative engine optimization) and AEO (answer engine optimization) audit asks a different one: when someone asks an AI assistant a question in my space, does my site get pulled into the answer, and do I get credited for it?

Those aren't the same question, even though I had almost always assumed they overlapped. That gap turned out to be the whole story.

The four dashboards I started from

I wanted hard data behind every finding, pulled from every source that matters and checked against what Google itself now says about AI search. So I used four sources:

1. Ahrefs Site Audit. The technical crawl: broken links, duplicate pages, title and meta problems, what is and isn't indexable.

2. Google Search Console. What Google actually does with my pages. Which ones it indexes, which it crawls and then ignores, and where they rank.

3. Bing Webmaster Tools, the AI Performance report. This one is badly underused. It shows which of my pages Copilot cites, on which questions, and how much of the answer comes from me.

4. Google's own recent guidance on AI search. Their public statements on what does and doesn't affect AI Overviews and AI Mode, so I would not burn time on tactics they have already said do nothing.

That last source saved me a weekend of wasted effort. Google has now said plainly that the AI-specific tricks people sell, the special files and the schema aimed at chatbots, do nothing for Google's AI answers. Being crawlable and indexed is the whole game there. Worth knowing before you build for it.

Then I handed all of it to Claude Code and had it run a deep audit against my actual codebase and my live site at the same time.

How I ran it with Claude Code

The setup is what makes it useful, so here is the shape of it.

I pointed Claude Code at the repository and the production site together, and asked it to do four things. Crawl the live pages the way an AI bot would see them. Read the code behind those pages to find the cause behind each problem rather than the symptom. Compare everything against the four dashboards and Google's guidance. Then argue against its own conclusions before reporting back. Across all of these steps I also ran a separate, independent quality-gate agent over the work.

About twenty minutes later I had the report. Here is what mattered in it.

What the audit found

The headline finding was uncomfortable and clarifying at the same time.

Google is crawling a large block of my pages and then choosing not to index them at all. That sits a step below ranking low. A page on page two is at least in the running; these never make it into the index to begin with. Search Console labels them "crawled, currently not indexed."

The same pages are the ones AI engines cite the most. Bing's data shows my site earned 536 Copilot citations in three months. On the question "side projects for experienced consultants," my pages make up 62.5 percent of all the sources Copilot cited. On "best developer marketing newsletters 2026," 57.8 percent. Those numbers come from the pages Google threw away.

So two systems look at the same content and reach opposite verdicts. Google sees thin pages and bins them. AI engines see useful answers and quote them. I had been optimizing for one referee while the other one was already paying me.

The cause was sitting in my own code. 126 of my 130 programmatically generated pages were built from a single shared template that optimized for the keyword and left the content thin. Two of them were word for word identical apart from a single phrase. That is the textbook definition of thin, mass-produced content, and Google's recent updates target exactly that. The audit traced it straight to the function generating the pages, with the evidence attached.

A few smaller findings stung because they were so basic. My sitemap, the file that tells search engines what exists, listed zero blog posts. My freshest and most genuinely written content was invisible to crawlers. Several titles ran too long and got chopped off mid word in the results. A handful of pages were stamped with a fake publish date.

And one thing I needed to hear: the technical plumbing I had been worrying about was fine. Pages render fully without JavaScript, AI crawlers get the same content a human does, the structured data is mostly sound. I had been ready to rebuild things that didn't need touching. The audit told me to stop and look somewhere else.

The trap I almost walked into

The obvious fix for thin pages is to hide them from search. Tag them "do not index" and move on. If I had done that on instinct, I would have pulled those pages out of Bing too, and killed the 536 citations that are currently my single best distribution channel.

So the plan goes the other way. I will rewrite the pages that have real demand or real citations until they earn their place back in Google, and remove only the genuine dead weight that no engine wants. The data decides which is which, one page at a time, instead of a gut call.

How to run this on your own site

You don't need my stack or my tools for this. The loop is the same for any site, and it has four steps.

1. Source. Gather your own data before you ask anyone for advice. Pull your Ahrefs crawl (or whatever you use), your Search Console pages report, and if you get any AI traffic at all, your Bing Webmaster AI Performance data. Add a link to the latest official guidance from Google on AI search, so the audit stands on facts and not blog rumor.

2. Action. Hand all of it to your coding or research assistant in one go. Ask it to compare your live pages against your real data, find the cause behind each problem rather than the symptom, and then challenge its own findings before it answers. Tell it directly: argue against yourself first.

3. Outcome. Read the findings as a map of where each engine agrees and disagrees about your site. The disagreements are where the opening is, the way Google and AI disagreed about mine.

4. Changes. Sort the fixes by effort against impact. Do the cheap, high impact ones first. Protect anything that is already earning attention. Remove only what no engine values. Then measure again in a few weeks, so you know what actually moved instead of guessing.

What I'm doing next

This week went past the audit itself. I implemented the first batch of fixes and pushed the code: making the blog visible in the sitemap, fixing the over-long titles, correcting the fake publish dates, and protecting every page that already has citations before touching anything else. The early read is encouraging, a little movement in the right direction already, though it is far too soon to trust a few days of data. So I am leaving the measurement running all week. Part 2 will show the real numbers, what happened in Search Console, in Bing, and in citations, with no rounding and no spin. If a change did nothing, I will tell you it did nothing.

I turned the audit into a checklist you can run

Enough of this is repeatable that I wrote the whole loop down: the four data sources to pull, the exact prompts I gave Claude Code, the instruction to make it argue with itself, and the effort-versus-impact sort at the end. It is a free markdown file, the same as always. Paste it into Claude, ChatGPT, or Gemini and have it run the same audit on your own site, step by step.

→ The GEO/AEO Audit Checklist

If you run a site, documentation, or a product where you want to be the source that engines rank and assistants cite, this is the one to grab. Run it this week while the idea is fresh. And if your own audit turns up something you can't make sense of, hit reply and tell me what you are seeing. I am glad to take a look and point you in the right direction.

What I'm Thinking About This Week

Whether the split I found is unique to my site or the normal state of things in 2026. One set of pages failing on Google while winning on AI engines is a strong signal that the two sources now reward different shapes of content on the same site. I am going to fix my worst offenders first and watch whether both dashboards move, or only one. If only one moves, that tells me the two games are further apart than I think.

Reply to this email with one of:

• Whether you have ever checked your AI citation count, and what the number was (the Bing webmaster console reports it for free, and most people have never looked)

• A page of yours that Google ignores but you suspect AI engines might be using

• The one finding from your own site you would least want an honest audit to surface

P.S. The toolkit, in the order I would use it: Visibility Audit → GEO/AEO Audit Checklist → How to Get Cited by AI → SEO Strategy Playbook → Title & Meta Audit → Sovereign Idea Workflow → Business Model Canvas. All free.

P.P.S. Know someone running a site who can't tell whether AI search is helping them or hurting them? This is the issue that shows them how to find out. Forward it, and they can subscribe at thesovereigntechnologist.com.

That’s all for this week.

See you next Thursday.