The phrase shows up on every agency homepage now. It rarely shows up with a definition behind it. Most of the sites flying that banner are running a 2019 template with a hero video, a parallax scroll, and a contact form. The H1 is keyword-stuffed. The body copy is six paragraphs of generalities. The schema is whatever Yoast or RankMath generated by default. The “AI” part is a chatbot widget bolted onto the lower-right corner that the owner has not looked at in seven months.
That site is not AI-optimized. It is barely web-optimized. It will rank on Google because Google has been generous about rewarding the basics. It will not get cited by Perplexity. It will not show up as a source in a ChatGPT answer. And in twenty-four months, when half of the buyer’s research happens before a Google query ever gets typed, that site is going to be invisible at the moment the decision is made.
The good news is the actual definition of “AI-optimized” is not vague. It is specific, narrow, and easy to audit. Most agencies sell against it because they cannot ship it.
What the AI platforms actually parse.
Perplexity, ChatGPT search, Claude, Gemini, and Google’s AI Overviews all rely on the same primitives. They read your HTML. They read your structured data. They read your sitemap. They follow your internal links. They evaluate the consistency between the schema you publish and the prose those schema entities reference. None of this is magic. All of it is decade-old web technology applied through a new evaluation lens.
What they do not do is reward decorative copywriting. The model is looking for a question and a fact-dense answer in close visual proximity, with enough surrounding context that the citation is defensible. If your page buries the answer to “how much does HVAC duct cleaning cost” inside three paragraphs of brand storytelling, the model is going to cite the page that put the number in the second sentence. A real number, with a range, with conditions. The page that says “starting at $325 for a single-system home, up to $580 with full register cleaning” beats every “Contact us for a custom quote” page every single time. That same answer-with-specifics structure is what unlocks rich-result eligibility in Google itself. Google’s Search Central documentation cites Nestlé measuring an 82% higher click-through rate on pages that show as rich results versus non-rich results, and Rotten Tomatoes seeing a 25% CTR lift across 100,000 structured-data pages. The page that gets cited by Perplexity is the same page that earns the structured-data CTR lift on the search engine that still owns the most query volume.
The platforms also differ in subtle ways. Perplexity weighs recent updates heavily and rewards sites with clear, schema-tagged FAQ pages. ChatGPT search treats entity consistency strictly: if your name, address, and phone number disagree between your site and Google Business Profile, you lose. Claude rewards authorship signals harder than the others and demotes anonymous pages. Gemini leans on the same signal stack as Google itself, which is good and bad: good if your traditional SEO foundations are solid, bad if you have been gaming any of them. We score all six platforms separately through our Lighthouse Local audit because the differences matter when you are debugging why one platform cites you and three do not.
The actual checklist.
We run every new client site through this. None of it is exotic. All of it is missing on most production sites we audit.
1. Answer-first information architecture. Every page on the site addresses one buyer query, named explicitly in the H1. The first 80 words contain the literal answer. The next section explains the conditions, edge cases, and trade-offs. The story comes last, not first. If your H1 is “Welcome to Our Plumbing Services” you have already lost the page. If your H1 is “How Much Does It Cost To Replace A Hot Water Heater In [City]?” and the next sentence is “$1,200 to $3,400 installed, depending on tank size and venting requirements,” you are in the citation pool.
2. Schema that matches the prose. If the page says “we serve Austin and Round Rock,” the LocalBusiness schema lists Austin and Round Rock in areaServed. If the page lists prices, the schema includes priceRange or Offer.price. If the page lists reviews, the schema includes aggregateRating with reviewCount matching the real review feed. The model checks for agreement between the structured data and the surrounding text and discounts pages where the two disagree. A schema field that contradicts the prose is worse than a missing field.
3. Real authorship and review signals. Author byline tied to a Person schema entity with a sameAs URL pointing to a LinkedIn, a credentialing body, or a published bio. Review schema with real review text exposed in the HTML, not the four-star aggregate rendered as a static image with no underlying data. The platforms have been burned enough times by fake aggregate ratings that they now want the underlying reviews accessible in the HTML. Sites with five hundred review badges and zero parseable review entities are getting filtered out of citation lists by mid-2026.
4. Internal linking by topic, not navigation. The site links from the HVAC service page to the HVAC pricing page to the HVAC FAQ page to the HVAC reviews page. Topic clusters, not breadcrumb-style “back to services” links. The model uses internal link density to figure out what the site is about. A site with eight service pages and one shared FAQ for the whole company looks like a generalist. A site with eight service pages and eight category-specific FAQs looks like an authority. We document the cluster work in detail under our content and editorial practice.
5. Page weight under 1 MB, TTI under 1 second. Crawlers throttle requests on slow sites. AI crawlers throttle harder than Googlebot, in our measurements, by a factor of about three. A page that takes 4 seconds to render gets crawled less often. A page crawled less often shows up in fewer AI answers. There is no exception for this. If your site is on a CMS that ships 380 KB of jQuery and tracking pixels before the H1 paints, the technical work has to come before the content work.
What “AI-optimized” is not.
The phrase is used to describe three things that have nothing to do with being cited by AI.
It is not a chatbot widget. A chatbot on your page does not help the AI platforms understand your site. The chatbot does not speak with the same model the buyer is using to find you. They are unrelated systems. A well-built AI agent on inbound is a separate and valuable layer, but it sits on top of an AI-optimized site, not in place of one.
It is not an AI-written blog. Pages drafted by an off-the-shelf model and pushed to production without editorial work read as generic to the same models that wrote them. The platforms can identify their own slop and they downrank it aggressively. The content that gets cited is the content that contains specifics no model could have generated without your team’s input.
It is not a redesign. Most “AI-optimized website” pitches we see in the wild are visual redesigns. New typography. New hero photo. A pricing comparison table. None of that touches the information architecture, the schema, or the citation logic. The buyer pays for a redesign, the agency ships it, the AI citation rate does not move, and everyone blames the platforms.
What we ship when we rebuild a site.
The deliverable is documented across our AI-optimized websites service, but the short version is straightforward.
A static, edge-served site with green Core Web Vitals on every page from launch. An answer-first content architecture mapped to the buyer queries that actually drive revenue. Schema deployed and entity-mapped so the citation lands on your business. Inbound chat and call routing layered at the edge so a visitor never waits. A handover that includes the code, the design system, any headless CMS we wired in, and the deployment infrastructure. You own all of it at engagement end.
We pair the build with continuous scoring inside Lighthouse Local across the six platforms, so you can watch citation rate move month over month against real prompts your buyers would actually type. For retainer clients, the site rebuild typically sits inside a broader engagement that also covers local SEO or national SEO, depending on the footprint.
Frequently Asked Questions.
How long until the citation rate moves after a rebuild?
Six to twelve weeks for measurable movement on the platforms that crawl most aggressively (Perplexity, ChatGPT search). Twelve to twenty-four weeks for the slower platforms (Claude, Gemini). The pattern we see is that citation rate moves first on schema-heavy questions and last on broad recommendation queries.
Can an existing site be retrofitted, or does it need a full rebuild?
Both work in different cases. A site already built on a static-first architecture (Astro, Next.js static export, Hugo) can usually be retrofitted with schema and content work alone. A site on a heavy CMS with custom theming usually rebuilds faster than it retrofits. The decision lives in the week-one audit.
Does AI-optimization come at the cost of Google ranking?
No. The work that gets you cited by the AI platforms is the same work that holds your Google rankings. They stack rather than trade. We have not seen a single case where a properly executed AI-optimization engagement cost a client Google rankings.
Is this just for content sites, or does it work for service-area businesses?
Both. Service-area businesses (plumbing, HVAC, roofing, electrical) benefit heavily because most of their competitors have ignored the work entirely. A well-optimized service-area site can pull citation rate from zero to thirty in a single quarter when the field is open.
The audit.
Run Lighthouse Local against your site. Score under 80 on the AI visibility category means the site is missing two or more of the items above. We can fix it, or you can fix it. Either way, fix it before you spend another quarter buying ads against a page no AI platform is going to cite.
