A growing share of buyers ask an assistant before they ask a search engine. “Who should I get to build a website in Toronto”, asked of ChatGPT, Perplexity, Copilot or Google’s AI Mode, returns a handful of named companies and a short justification for each — and the businesses not in that handful never learn they were considered.
Every agency in this market now sells optimisation for it, under a rotating set of acronyms. Most of what is being sold does very little. The parts that work are unglamorous.
How an answer actually gets built
An assistant answering a commercial question does roughly three things. It retrieves a set of documents, either from a search index or from its own crawl. It extracts passages from them. It composes an answer from those passages and attributes some of them.
Each step is a filter. A page that cannot be fetched never enters the set. A page that is fetched but whose relevant sentence cannot survive being separated from the page around it never becomes a passage. A passage that does not clearly identify who it is about gets used without attribution — which is the worst outcome, because the work was done and the credit went nowhere.
Optimising for this means making a page survive all three steps. Not one.
Step one: be retrievable
Nothing else matters if the page cannot be fetched.
Server-rendered HTML. Content that only exists after JavaScript executes is invisible to a large number of crawlers. Some render, many do not, and none guarantee it.
Crawler access. AI crawlers are separate user agents from Googlebot — GPTBot, OAI-SearchBot,
ClaudeBot, PerplexityBot, and others. Blocking them in robots.txt removes the site from
their answers entirely. That is a legitimate business decision for a publisher protecting
its content; for a service business trying to be recommended, it is self-harm, and it is
frequently done by accident when a security plugin adds a blanket bot rule.
Speed and stability. Crawlers give up on slow responses. A site that times out under load is a site that is not in the index.
The way to know rather than assume is to look at server or CDN logs and see which of these agents actually fetched the site, and how often. That report takes an hour to build and settles the question permanently.
Step two: be quotable
This is the part almost nobody does, and it is where the leverage is.
A retriever cuts pages into chunks of roughly 130 to 170 words, usually around a heading. Whatever sits under that heading has to make sense on its own, because that is the form it arrives in.
Three rules follow directly:
Never open a section on a pronoun. “It handles Canadian tax by province” is useless once extracted — the reader has no idea what “it” was. “Shopify handles Canadian tax by province” survives the journey. This one change does more than most of what is sold as AI optimisation.
Name the subject inside the passage. The company name, the city, the product. A page can be entirely about your business without naming it anywhere below the logo, and an extracted passage from that page is unattributable.
Avoid relative time near facts. A sentence saying a business “currently” charges a given amount is wrong the moment it is quoted six months later, while the same sentence written as a plain statement of what a service starts at stays true. Dates that matter should be absolute.
Answering real questions in plain declarative sentences — with the numbers in them — outperforms fluent marketing prose, because marketing prose contains almost nothing that can be lifted and used.
Step three: be identifiable as an entity
An assistant composing an answer has to be confident about which real-world thing a page refers to before it will attach a name to it.
Structured data does this job when it is done properly, which mostly means using stable
@id anchors so every page agrees about what the organisation is, rather than each page
declaring an unconnected Organization node. Consistency of the basics — the same business
name, phone number and description across the website, the Google Business Profile and any
directory listings — does the rest. Contradictory details across sources are a reason for a
model to hedge, and hedging means not naming you.
Third-party corroboration matters more here than in classic search. A business that appears consistently on its own site, its Google Business Profile, and a couple of credible industry directories is easier to be confident about than one that exists in a single place.
What to ignore
llms.txt as a growth lever. Google Search does not use it for ranking or AI Overviews, most published files are never fetched, and no measured citation effect exists. It costs nothing to publish and some retrieval workflows respect it, so this site publishes one — but anybody selling it as an AI-search strategy is selling a file, not a result.
Keyword density, entity stuffing, and “GEO scores”. No answer engine publishes a scoring model, and vendors offering a number for how AI-ready a page is have invented the number.
Guaranteed citation. Answer engines return different results to the same prompt on different runs. Anybody promising to place you in an AI answer is promising something they cannot observe, let alone control.
Measuring it, which almost nobody does
The reason this field is full of unfalsifiable advice is that most agencies never check.
Measurement is not difficult. Write down the questions a real buyer would ask — “who should I hire to build an online store in Toronto”, “how much does a website cost in Toronto” — put them to each engine on a schedule, and record two things per run: whether the business was named, and whether it was cited with a link. Track the trend and note which competitors appear alongside.
InceptiCode does this for itself. The prompt set, the run history and the results live in the site repository, alongside a report of which AI crawlers actually fetched the site and how often. That is a modest, honest instrument, and it is more than a scoring dashboard provides, because it measures the outcome rather than a proxy for it.
The summary
Be fetchable. Write passages that survive being cut out of the page. Be unambiguous about who you are, in markup and in words, everywhere you appear. Then measure whether any of it worked, and be suspicious of anybody who does not.