We Audited Our Own Site for AI Search — Here's What Actually Mattered

The short version: we spent a week optimizing our own website for AI search — schema, crawler policy, answer-first content, the lot. The single highest-value thing we found had nothing to do with any of that. It was that our website contradicted itself, and we had never noticed.

Here's the full account, including what we'd tell you to do first if you're thinking about the same thing.

What AI search optimization actually is

Two terms get used for this, and they're worth separating.

Answer Engine Optimization (AEO) is structuring your content so an AI assistant can extract a clean, accurate answer from it. Traditional SEO competes for a position in a list of links. AEO competes to be the sentence the assistant actually quotes.

Generative Engine Optimization (GEO) is the broader practice of getting your brand named inside AI-generated answers — ChatGPT, Claude, Perplexity, Google AI Overviews. The unit of success isn't a ranking. It's a mention.

Both matter more each month, because the behaviour has already shifted. People who used to search "best web design agency near me" now ask an assistant the same question and take the three names it gives them.

The thing that mattered most (and it isn't technical)

When we went through our own domain systematically, we found this:

Our website said we had 15+ years of experience in four places, and 10+ years in two others. It promised a 24-hour response time in seven places, and 48 hours in one.

Now picture what happens when someone asks an assistant how experienced we are. The model reads our domain, finds two different answers, and does one of two things: it picks one at random and may state it wrongly, or it treats the source as unreliable and reaches for a competitor whose facts line up.

Neither outcome is good, and neither shows up in any analytics dashboard. You simply don't get mentioned, and you never find out why.

Fixing it wasn't clever work. It was reading our own pages and making the numbers agree. But it was the highest- value item on the entire project, and it cost nothing but attention.

What else we changed

Entity definition

We had a bare business listing with a single social link, which meant our LinkedIn, Medium, Facebook and Google profiles read as unrelated sources rather than one company. We declared all of them together with a stable identifier, added both founders as named people, and listed our services and topic areas explicitly. This is how a model verifies you're a real, consistent organisation rather than a name on a page.

Answer-first passages

Marketing copy is written to build toward a point. AI extraction wants the point first. A hero line like "Do more with less" reads well to someone who already knows what the page is about, and gives a model nothing to lift. We added a direct two- or three-sentence answer near the top of every service and industry page — written so it still makes sense pulled completely out of context.

Definitional content

Pages that define a term get cited far more often than pages that sell a service, because they answer the question being asked rather than arguing against it. We published a plain guide to what AI automation is, including what it isn't.

Crawler policy

We named the assistant crawlers explicitly — GPTBot, ClaudeBot, PerplexityBot, Google-Extended and others. Our existing rules already allowed them, but naming them is the convention and it survives any future tightening of the wildcard rule.

What we'd skip

There's a lot of noise in this space right now. Two things we'd deprioritise:

llms.txt files. Google has said plainly that no special file is needed for its AI features, and adoption elsewhere is thin. We keep one because it's harmless and doubles as a plain-language summary of what we do — but it is not the lever that gets you cited, and treating it as a magic switch will waste your week.

Writing for models instead of people. Keyword-stuffed, thin content performs worse in AI answers, not better. Every credible finding in this area points the same direction: genuinely useful content, clearly structured, wins.

The half most people skip

Everything above is on-site work, and on-site work has a ceiling. Schema and structure help an assistant understand who you are. They don't make it recommend you.

Recommendation comes from being mentioned in places the model already trusts — industry directories, review platforms, roundup articles, community discussions. If your name only appears on your own domain, you're a well-organised stranger.

That's the harder half, it takes months rather than days, and it's where most of the remaining upside sits. It's also, conveniently, the same work that builds traditional SEO authority. One effort, two channels.

How to audit your own site this week

1. Find your contradictions. List your key claims — years in business, response time, pricing, service area, team size. Search your own domain for each one. Count the versions. Fix them.

2. Read your top pages cold. Take the first three sentences of each key page and ask whether they answer an obvious question on their own. If they only make sense in context, rewrite them so they don't have to.

3. Test how you currently appear. Ask ChatGPT, Claude and Perplexity the questions your customers would ask — "best web design company in [your city]", "who can build a website for my small business". Note whether you appear at all. Repeat monthly. That's your baseline.

4. Claim your listings. Business directories relevant to your industry, your Google Business Profile, review platforms. This is the part that turns understanding into recommendation.

Frequently Asked Questions

What is the difference between SEO and AEO?
SEO competes for a position in a list of links. AEO competes to be the sentence an AI assistant lifts out and uses as its answer. The same foundations power both, but AEO rewards answering directly in the first sentence rather than building to it.

Do I need an llms.txt file for AI search?
Not as a priority. Google has stated no special file is required for its AI features. It's harmless and can serve as a plain summary of your business, but consistent facts, clear structure and third-party mentions matter far more.

How do I check if my website contradicts itself?
Pick your key claims and search your own domain for each one using site:yourdomain.com followed by the claim. Count how many different versions exist. Most businesses find more than one.

How long does AI search optimization take to work?
On-site fixes take effect as soon as pages are recrawled — days to a few weeks. Being named in AI answers depends on third-party mentions, which build over three to six months.

The takeaway

We went in expecting the win to be technical. It wasn't. The most valuable hour of the whole project was spent reading our own website carefully enough to notice it was telling two different stories.

Before you touch schema, before you write a single file for the crawlers: open your site and check whether it agrees with itself. Most don't. Ours didn't.