For twenty years, the goal of a business website was to rank on a page of search results and get the click. That’s quietly changing. More and more, a potential customer doesn’t see a page of ten links; they ask an assistant a question and get one answer. Google’s AI Mode search just passed a billion users. People ask ChatGPT for recommendations constantly. And Apple just announced that its new Siri is powered by Google’s Gemini, which means the assistant in a billion pockets is about to get a lot more willing to answer “who’s a good web studio in Toronto?”

The new question isn’t “do I rank?” It’s “when an AI answers a question my customer asks, does it mention me, correctly?”

That capability has a name people are settling on, answer engine optimization or AEO, but the work is less exotic than the acronym. It rewards the same things a well-built website has always needed, just more strictly, because now a machine has to read and trust your site well enough to vouch for you in an answer.

What “AI-readable” means

An AI assistant recommending your business is doing three things: finding your site, understanding what you do, and trusting it enough to repeat it. Here’s what each one needs.

1. Structured data, so machines know what you are

Humans infer from layout that you’re a Toronto dental clinic open until 6pm. A machine needs it spelled out in a format it parses without guessing: JSON-LD structured data (LocalBusiness, Organization, Product, FAQPage, and so on). This is the single highest-leverage, most-overlooked thing on most small business sites. It’s invisible to visitors and decisive for assistants.

2. Content shaped like answers

Assistants extract answers. A page that buries “how much does a kitchen reno cost” under three paragraphs of brand voice gives them nothing to lift. Pages with clear questions as headings and concise, honest answers underneath get quoted. This isn’t keyword stuffing; it’s writing the way people actually ask, and answering plainly.

3. Speed and clean markup, so it can be read at all

If your site is a slow, JavaScript-heavy template where the real content only appears after a spinner, crawlers and agents frequently get an empty page. Fast-loading, server-rendered, semantic HTML isn’t just a Core Web Vitals nicety anymore; it’s the difference between being legible to an assistant and being skipped. (This is a big reason we build on Astro and render content on the server.)

4. Consistent, accurate facts everywhere

If your phone number, hours, or address differ between your site, Google Business Profile, and old directory listings, an assistant has no way to know which is right, so it may pick the wrong one or omit you to be safe. Consistency is trust.

Why you can’t really game this

The part we like is the honest part: there’s no trick. You can’t buy your way into an AI answer the way people once bought backlinks. These systems are explicitly built to surface sources that are accurate, well-structured, fast, and useful. The “optimization” is just building a good website properly, which is what we’ve argued small businesses deserve all along. The AI shift didn’t invent new rules; it raised the penalty for ignoring the old ones.

The flip side: if your site is a template-mill build with thin content, no structured data, and a slow load, you’re now invisible in two places at once: the old search results and the new AI answers.

A five-minute test you can run today

  1. Open Google’s AI Mode (and ChatGPT) and ask the questions your customers ask: “best [your service] in [your city],” “who does [specific thing] near me,” “is [your business] any good.”
  2. Note whether you appear, and whether what’s said is correct.
  3. Ask an assistant to summarize your homepage. If it struggles or gets your services wrong, that’s a structure-and-clarity problem, not a mystery.

What you find is a surprisingly honest report card. Most small business sites we audit fail it for boring, fixable reasons: no structured data, content that doesn’t answer real questions, and a slow build.

What we do about it

Every site we build ships with this baked in: JSON-LD for the whole business, answer-shaped content, server-rendered speed, accessible semantic markup, and consistent business data. Not because it’s a 2026 buzzword, but because it’s what makes a site work for the people and the assistants now standing between you and them.

If you want to know how your current site scores, and what it would take to show up when a customer asks an AI about your line of work, send us the URL and we’ll tell you straight.

Update, August 21, 2026: a correction, and what Google actually says now

When this was written in June, nobody had official guidance on any of this, so the advice above was assembled from how these systems behave. Google has since published a formal document, its guide to optimizing for generative AI features on Search, last updated July 10, 2026. One part of it contradicts the strongest claim in this post, and we would rather correct it than leave it.

Structured data is not required. Google’s wording is that structured data is not required for generative AI search and there is no special schema.org markup you need to add. Section 1 above calls JSON-LD “the single highest-leverage, most-overlooked thing on most small business sites” for being recommended by an assistant. That was overstated. Structured data is still worth shipping, for rich results in classic Search, for Google Business Profile and Merchant Center consistency, and for the many non-Google systems that read it, and every site we build still ships it. But it is not the lever on AI Overviews and AI Mode that this post implied.

The same document rules out a list of tactics being sold as AI-search packages. You do not need new machine-readable files, AI text files, markup, or Markdown, and it names llms.txt specifically. You do not need to chunk content into small pieces. You do not need to write in a special way for generative AI, or make a page per query variation, or chase mentions across other sites. We publish an llms.txt on this site and we are keeping it, because it costs nothing and other systems may use it, but we are not going to pretend it is doing anything for Google’s AI features.

What Google does say matters: unique content with a real perspective rather than commodity content that recycles what others have said, pages that are indexable and technically sound, semantic HTML, good page experience, and, for local and commerce, Merchant Center and a maintained Business Profile. That is sections 2, 3, and 4 above, unchanged, and it is the same conclusion from a different direction.

Two other things have changed since June. Search Console now has a generative AI performance report and a control that excludes your site from AI Overviews, AI Mode, and Discover’s AI features entirely. The report shows impressions only, with no clicks and no click-through rate, and the control is worth understanding before somebody with access to your property flips it. We covered both in the write-up on Google’s AI search controls.

And on the billion-user figure in the opening: it is Google’s own, and it is about people rather than searches. SparkToro’s analysis of Similarweb clickstream data for January through April 2026 found that 0.34% of US Google searches transitioned into AI Mode. Both numbers are true. A great many people have tried AI Mode; the overwhelming majority of searches still go through the ordinary results page. Build for the assistants, but do not let anyone tell you classic search stopped mattering this year.

Update, September 11, 2026: your website is now somebody’s customer service script

Everything above is about being found. There is a second consumer of your website copy now, and it changes what “accurate” costs you when you get it wrong.

Meta’s Business Agent, which Meta says more than a million businesses are already running on WhatsApp and Messenger, builds its knowledge of your business by learning automatically from your Facebook Page, your website content, your catalogue and your past conversations. So does every competing agent in this category. That means the paragraph on your services page that has been slightly out of date since you raised prices in March is no longer just a bad impression. It is the answer a customer gets at 11pm, in your business’s name, with an “AI” label on it, from something that sounds completely certain.

This is a different and sharper test than the one in the rest of this post. Being recommended by an assistant rewards depth, perspective and technical soundness. Being spoken for by an agent rewards something narrower and duller: hours, pricing, service area, what you do not do, and what happens if a customer wants a refund, all stated once, unambiguously, and consistent with every other place you state them. Contradictions used to cost you a confused reader. Now they get resolved by a machine picking one, and it will not tell you which.

The cheapest version of this audit is to read your own site as a stranger looking for a reason to buy, and write down every question you cannot answer from the page in front of you. That list is what your agent, and everybody else’s, will be guessing at. We went through the economics and the eligibility rules of the Meta product in the post on Business Agent billing; this part applies whether or not you ever turn one on, because your competitors’ agents and the general-purpose assistants are reading the same page.

Update, September 13, 2026: the third consumer, and it is not reading your copy at all

The post above has two audiences for your website: the assistant deciding whether to recommend you, and the agent reading your pages to speak for you. There is now a third, and it is a different kind of reader. It does not parse your prose. It calls functions.

WebMCP is a draft standard from the W3C Web Machine Learning Community Group that lets a page register named tools an AI agent can invoke directly, through document.modelContext.registerTool. Instead of an agent reading your booking page and guessing which button to press, the page hands it a check_availability function with a schema. Chrome has had it in origin trial since Chrome 149 in June, Edge since Edge 150, and on August 25 OpenAI made ChatGPT’s desktop browser able to call these tools on any page that offers them.

Two things about it are worth your attention and one is worth ignoring.

Worth ignoring: it has nothing to do with being found. WebMCP only matters once an agent is already on your page. It does not affect ranking, citation, or whether an assistant recommends you. Everything in the rest of this post still does that job, and no amount of tool registration substitutes for any of it.

Worth your attention if you sell online. On August 5 Shopify switched these tools on across every Liquid storefront, with nothing to install and no admin toggle. Ten of them, including one that reads your policy pages and FAQs and hands the text to an agent to paraphrase. That makes the update above literal rather than figurative: your returns policy is now an API response with a schema, and the stale sentence in it gets read out as your position. We checked sixteen live storefronts and wrote up exactly what is exposed in the post on Shopify’s WebMCP tools.

Worth your attention if you are being sold on it. Do not add tools to your own site this quarter, and especially not by installing a plugin. The API’s entry point moved from navigator.modelContext to document.modelContext on July 21, Chrome deprecated the old name in Chrome 150, Chrome’s origin trial for the whole thing runs out at Chrome 156 on October 20, and Chrome Platform Status still lists the feature as Proposed with no committed shipping milestone. WebKit’s public position on the proposal is oppose and Mozilla’s is neutral, so Safari and Firefox are not part of this yet. The cost of waiting until Chrome decides is roughly nothing. The cost of building against an API that has already moved once this summer is not nothing.