What Your Website Has to Do Before AI Agents Will Buy From It

Your website now has two readers, and only one of them clicks. The other is a crawler or an assistant pulling a quotable answer out of your page so it can hand a buyer a shortlist, and most redesign briefs still do not account for it. That is the gap behind the pattern we keep seeing: traffic holds, rankings hold, pipeline thins.

The fix is not a plugin. It is a build decision. Structure the site so a machine can extract, verify and cite what you do, who it is for, what it costs to engage and why you can be trusted, then let design do the human job of making a buyer believe it.

Your new site has two readers

A human lands on your homepage and skims for trust. Big type, motion, a few proof points, a face. They make a gut call in seconds, then go looking for evidence.

A machine does something else entirely. It fetches the HTML, parses what is there in text, and moves on. It does not watch your hero video. It does not wait for a carousel. It does not infer your pricing model from a beautiful diagram. It needs the answer written down, in the markup, in plain language, attached to an entity it can identify.

Both readers are looking for the same four things. Only one of them can read between the lines.

What an AI engine can actually pull off your page

The practical list is short and unglamorous.

• Structured entity data. Organization, Product, Service, Person, FAQ, Offer. Schema is how you tell a machine what kind of thing it is looking at instead of hoping it guesses.
• Clear service and product definitions in text. One page per service, written as a definition first and a pitch second. "We do X for Y companies, here is what is included, here is what it costs to start."
• Specs and comparisons as text. Tables and lists in HTML, not screenshots of tables, not PDFs, not a slide exported as a JPEG.
• FAQ architecture. Real buyer questions as headings, each answered in a self-contained paragraph that survives being lifted out of the page.
• Named authors and verifiable claims. A person attached to the content, a number attached to the claim, a source attached to the number.

That last one is doing more work than most teams realize. Machines do not reward confidence, they reward checkability.

The build decisions that make a site invisible to machines

Five recur in almost every audit.

1. Client-side rendering. This is the big one. In Vercel and MERJ's December 2024 analysis of AI crawler behavior, ChatGPT's crawlers fetched JavaScript files in 11.50% of requests and Claude in 23.84%, and neither executes them. Their conclusion was blunt: these crawlers cannot read client-side rendered content. Gemini and AppleBot are the exceptions, Gemini because it inherits Google's rendering infrastructure. If your copy, your pricing, your FAQ or your product specs only appear after JavaScript runs, a large share of the machines deciding shortlists see an empty page.

2. The hero video carries the only real messaging. Cinematic opens are good brand work. They are also invisible as text. If the clearest statement of what you do exists in a voiceover, it does not exist.

3. Image-locked and PDF-locked content. Pricing in a graphic. Specs in a downloadable sheet. Comparison grids exported as images. In the same December 2024 dataset, images accounted for 35.17% of Claude's fetches on nextjs.org, while ChatGPT's fetches there were 57.70% HTML. Crawlers are pulling your images. They are not reading the numbers inside them.

4. Duplicate thin service pages. Twelve near-identical pages built to chase keyword variants tell a machine you have one vague offering, not twelve capabilities. Fewer, deeper, differentiated pages extract better.

5. No schema at all. Still the most common finding, still the cheapest to fix.

Replatforming is the riskiest moment for your AI visibility

Here is the part almost nobody puts in a redesign plan.

AI crawlers waste an enormous share of their budget on URLs that no longer exist. In the Vercel and MERJ data, 34.82% of ChatGPT's fetches hit 404s and 34.16% of Claude's did, with another 14.36% of ChatGPT's fetches going to redirects. Googlebot, by comparison, hit 404s on 8.22% of fetches and redirects on 1.49%. These crawlers are considerably less efficient at discovering what moved.

Translate that to a launch. You change your URL structure, your old service pages disappear, and for a stretch after go-live a meaningful portion of AI crawl activity on your domain is spent hitting pages that are gone. Meanwhile the answer being served to buyers is built from whatever was cached before.

So the migration plan matters as much as the design. One to one redirects for every URL that ever earned a citation, no redirect chains, a clean sitemap published at launch, and no "we will clean up the 404s next quarter."

The pre-launch test

Before you sign off a build, run one test. Strip the page to its text. No images, no video, no scripts. Then ask whether a machine could answer these four questions from what is left:

1. What does this company do? Named, specific, in the first screen of text.
2. Who is it for? Segment, size, situation. "Growth-minded brands" is not an answer. "Shopify brands doing eight figures with an in-house marketing lead" is.
3. What does it cost to engage? A range, a starting point, a model. If you will not publish a number, publish the shape of the number.
4. Why should it be trusted? Named people, named clients, specific outcomes, dated claims.

If the answer to any of those lives only in a visual, it is not on your site as far as an agent is concerned. It is far cheaper to argue about it in wireframes than in a retro.

Stop measuring traffic

The uncomfortable truth about AI visibility is that the volume is small. In Amsive's analysis of 54 sites, organic search drove about 31.9% of sessions while LLM referrals drove 0.24%. You are not replacing your traffic stream. You are optimizing a tiny, high-intent one.

What makes it worth the engineering is what those visitors do. Ahrefs reported from their own site in June 2025 that AI-referred visitors were 0.5% of its traffic but drove 12.1% of signups, converting 23x better than traditional organic visitors. Semrush research estimates AI search visitors convert 4.4x better than traditional organic search visitors. Worth noting the other side: Amsive's 54-site study found no significant overall difference, a mean conversion rate of 4.60% for organic against 4.87% for LLM traffic, p = 0.794. The split underneath it matters more than the null result: B2B sites favored LLM traffic, 2.17% against 1.16%, while B2C was near identical at 6.58% against 6.78%. Single-site and vendor-reported numbers are directional, not law.

Which is why traffic is the wrong success metric for a new site. Three better ones:

• Citation rate. How often you appear in AI answers to the questions your buyers actually ask. Tracked by prompt, by engine, over time.
• Assisted pipeline. Deals where an AI assistant touched the research path, captured at the point of conversion by asking.
• High-intent page conversion. Pricing, comparison and service pages. Fewer sessions converting harder is a win, and a sessions-only dashboard will report it as a loss.

And do not assume your Google rankings carry over. The overlap between what ranks first on Google and what AI engines actually cite has collapsed, and the signals that earn a citation are not the signals that earn a ranking. Treat ranking and being cited as two different jobs.

The short version

Design for the human. Structure for the machine. Server-render the substance, define your services in text, publish the specs and the prices as words, attach names to claims, and treat the migration as the highest-risk moment in the project.

We build sites this way because we run AI visibility as a discipline, not a bolt-on, and BeFound AI exists precisely to measure what the new scoreboard is reading. If you are scoping a build or a replatform and want the machine-readability requirements written into the brief before anyone opens Figma, that is a conversation worth having early.

Frequently asked questions

Does schema markup alone make our site readable to AI engines?

No. Schema tells a machine what kind of thing it is looking at, but it cannot supply content that is not there. If your pricing lives in an image, your positioning lives in a hero video, or your copy only renders after JavaScript runs, schema has nothing to describe. Fix the content layer first, then mark it up.

We rank well on Google. Doesn't that mean AI assistants will cite us?

Not reliably. Ahrefs found that only 12% of AI-cited URLs rank in Google's top 10 for the original prompt, with roughly 11% average overlap across assistants. Ranking and being cited run on different signals, so treat them as two separate jobs.

Is client-side rendering really a problem?

For most AI crawlers, yes. Vercel and MERJ's December 2024 analysis found GPTBot fetched JavaScript in 11.50% of requests and Claude in 23.84%, and neither executes it, meaning they cannot read client-side rendered content. Gemini is the exception because it uses Google's rendering infrastructure. Anything that matters should be server-rendered in the HTML.

AI referral traffic is tiny. Why invest in this at all?

Because the volume and the value are different things. One analysis of 54 sites put LLM referrals at 0.24% of sessions versus about 31.9% from organic search, but Ahrefs reported AI-referred visitors at 0.5% of sessions driving 12.1% of signups on their own site, and Semrush put the differential at 4.4x by conversion rate. Amsive found no significant difference, so the range is wide. You are optimizing a small, high-intent stream, not replacing your traffic.

What is the riskiest part of a redesign for AI visibility?

The URL migration. AI crawlers are far less efficient than Googlebot at finding what moved: in the Vercel and MERJ data, 34.82% of ChatGPT fetches and 34.16% of Claude fetches hit 404s, versus 8.22% for Googlebot. Plan one to one redirects for every URL that ever earned a citation, avoid redirect chains, and publish a clean sitemap at launch.