The 30-second version
People are asking ChatGPT and Google’s AI Overviews the questions they used to type into a search box, and businesses are starting to get customers who say “the AI recommended you.” So we did what we always do with a new channel: experimented on our own site first. We added an llms.txt file, tightened our schema so every page describes the same entities the same way, restructured headings into questions with direct answers underneath, added TL;DR blocks a language model can lift verbatim, and made bylines and dates explicit. Some of this is probably load-bearing, some is probably harmless overkill, and anyone who tells you they know exactly which is which is selling something. Here’s what we changed, why, and where our honest uncertainty sits.
The mental shift: you’re writing for a quoter, not a ranker
Classic SEO is a ranking contest: ten blue links, fight for position. AI search is a citation contest: the model composes one answer and decides whose content is worth quoting or naming inside it. That changes what “optimized” means. A ranker rewards a page that satisfies a query. A quoter rewards a passage it can extract cleanly: a direct claim, clearly attributed, in text that doesn’t require the surrounding page to make sense.
Almost everything we changed follows from that one idea. The good news is that none of it conflicts with normal SEO. Everything below makes the site better for regular Google too, which is your insurance policy if the AI channel evolves out from under us. That continuity is why we treat this as part of SEO, not a separate religion.
What we actually changed, piece by piece
llms.txt. A plain-text file at the site root that tells language models what the site is, who’s behind it, and where the substantive pages live. It’s an emerging convention, not a standard with guaranteed adoption, and we’re clear-eyed about that. We added it because it costs an hour, it can’t hurt, and if crawlers do lean on it, we’d rather be early. One rule we hold ourselves to: it claims only what the site’s pages actually corroborate. An llms.txt full of assertions your content doesn’t back up is asking a fact-checking machine to catch you exaggerating.
Entity-consistent schema. Language models assemble a picture of who you are from every mention across your site and the web. We rebuilt our structured data so there is exactly one Organization and one Person, each with a stable @id, referenced identically from every page: same name, same address, same author. Before, our pages described the business in slightly different ways, which is the digital equivalent of giving a different spelling of your name at every appointment. Consistency is how a machine becomes confident it’s talking about one specific entity, and confidence is what gets you named in an answer.
Question-shaped headings with direct answers. We rewrote headings to match how questions are actually asked, and made the first sentence under each heading answer the question outright. No warm-up paragraph, no “in today’s digital landscape.” Claim first, elaboration after. If you read our recent posts, you can see the pattern: every section leads with its conclusion.
TL;DR blocks. The “30-second version” at the top of every post exists partly for human skimmers and partly because a self-contained summary paragraph is the easiest thing on a page for a model to quote accurately. If something is going to be extracted, we’d rather write the extract ourselves.
Bylines, dates, and semantic HTML. A named author, a visible publish date, an honest dateModified, and clean article markup: heading hierarchy that means something, real lists, proper article tags. Models weigh provenance signals when deciding what’s trustworthy, and clean structure makes extraction less error-prone. This also stacks with site speed and clean output, which is a big part of why we build sites the way we do: a fast, plainly structured HTML page is easy for every crawler on earth, including the new ones.
What seems to matter most, and what’s snake oil
If I had to rank the levers by conviction: first, actually answering questions directly in your content, because that’s visible in how AI answers get assembled from sources. Second, entity consistency, because getting recommended by name requires the model to be sure who you are. Third, provenance signals like bylines and dates. Fourth, structural conveniences like llms.txt and TL;DR blocks: cheap, plausible, unproven.
Meanwhile the snake oil has arrived on schedule. “AI SEO packages” guaranteeing ChatGPT recommendations. Tools that promise to inject you into training data. Agencies certifying themselves in a discipline that’s eighteen months old. Nobody can guarantee an AI recommendation, the same way nobody could ever guarantee a ranking, and the recommendation engines are shifting under everyone’s feet monthly. What’s durable is the boring stuff underneath: being a real entity, with consistent facts, publishing genuinely useful answers. That was the winning strategy before AI search and it’s suspiciously identical now.
How we’re measuring, and what we honestly don’t know
We watch three things: referral traffic from AI surfaces, whether our pages appear as citations in AI Overviews for queries we care about, and the crude but useful habit of asking the assistants our customers’ questions and seeing who gets named. We also ask new leads how they found us, which catches the “ChatGPT told me” cases analytics can’t see.
What we don’t know is real: how much weight any single change carries, how stable citations are week to week, whether llms.txt will matter at all in a year. AI search is young and everyone measuring it honestly says the same thing. Our position is to make the changes that are cheap, aligned with good SEO anyway, and reversible, and to keep watching. That’s not a thrilling pitch, but it’s the true one.
Common questions
Is AI search actually sending businesses customers yet? Yes, in growing but modest volume, and it’s often invisible in analytics because the customer arrives by typing your name. Ask every new lead how they found you; that question now has a new possible answer.
Do I need to do anything if my SEO is already good? Less than the hype suggests. Direct answers, consistent entity data, and clean markup cover most of it. If your content already answers real questions plainly, you’re most of the way there.
Will this replace regular SEO? They’re converging, not replacing. AI answers are assembled largely from content that already ranks and is easy to extract from. Feed both with the same work.
Can I just block the AI crawlers instead? You can, and for some publishers that’s a rational choice. For a local service business, being absent from the place customers ask for recommendations is a strange thing to volunteer for. The technical details of access are in our AEO checklist.