AI Search Is Answering Your Customers. Here Is How to Stay in the Answer

A growing share of the questions your customers ask never reach a list of blue links. They get answered in an AI overview, in a chat assistant, or inside a tool that quietly called a search API on the user's behalf. The click you used to earn is now sometimes a citation, and sometimes nothing at all. That is uncomfortable, and it is survivable if you understand what these systems actually reward.

What the assistants are actually doing

Strip away the branding and the pattern is the same everywhere: the system runs one or more searches, fetches a handful of pages, extracts the passages that look like they answer the question, and writes a short synthesis with links. Your page is competing to be one of the passages that gets pulled, not to be the destination. That is a different game from ranking, though the inputs overlap heavily.

• Retrieval still starts with a conventional index, so classic technical SEO remains the entry ticket.
• Extraction favours pages where the answer sits in one self-contained passage rather than spread across a narrative.
• Synthesis favours sources that state something specific and checkable over sources that hedge.
• Citation favours pages the system can attribute cleanly: clear authorship, a visible date, a stable URL.
• Nothing here rewards length on its own. A 3,000-word page with the answer buried in paragraph nine loses to a 600-word page that answers in the first line.

Write the answer first and the context after

This is the single content change that matters most. If someone asks what a service costs, the number or the range belongs in the first two sentences under the heading, not in a paragraph that builds towards it. Extraction pulls passages, and a passage that only makes sense after four paragraphs of setup will not be chosen. Good technical documentation has always worked this way; marketing copy usually has not.

Then be specific enough to be quotable. Numbers, dates, named constraints and concrete steps survive extraction; adjectives do not. A sentence like "a marketing site of eight to twelve pages usually takes three to five weeks" can be cited. "We deliver fast, high-quality websites" cannot, because it contains nothing a system can attach to your name and nothing a reader can check.

Vague marketing copy was always weak. AI search made it invisible.

Structure the page so a machine can parse it

• One question per heading, phrased the way a person would ask it out loud.
• A direct answer in the first sentence under each heading, before any qualification.
• Structured data where it genuinely fits: FAQ, Product, Article, LocalBusiness.
• A visible published or updated date, so a system choosing between two sources can prefer the current one.
• Facts in text, not only inside an image, a PDF or a chart with no caption.

Check that the AI crawlers are allowed in, and decide which ones

Several separate crawlers are involved, and they are controlled separately, so the first job is to read your own robots.txt rather than assume. Search crawling, model training and answer-time fetching are distinct activities run by distinct user agents. Blocking the main search crawler removes you from search results and from the overviews built on them, which is almost never what a business wants. Training-specific controls are a different toggle and do not remove you from search.

• Know which user agents you allow, by name, and why. Most sites should default to allowing search crawlers.
• Training opt-outs and search crawling are separate controls, so check each rather than assuming one covers the other.
• Answer-time fetchers, which retrieve a page when a user asks a question right now, are the ones that decide whether you can be cited today.
• Rate-limit rather than block if crawler traffic is loading your server; a blanket disallow is a blunt fix with a large cost.
• The llms.txt convention is a proposal, not something anyone is obliged to honour. Adding one is cheap; treating it as a ranking mechanism is wishful.

Check the crawl side properly while you are in there. Blocked resources, slow responses, broken canonicals and orphan pages hurt in exactly the same way they always did, and our technical SEO audit checklist covers the pass we run for this.

If the content needs JavaScript, assume it will not be read

Answer-time fetchers are generally simpler than a full search crawler: many of them request the HTML, read what is there, and move on without executing scripts or waiting for a client-side data fetch. A page whose text arrives only after hydration may rank acceptably in traditional search, where rendering happens on a delay, and still be invisible to an assistant fetching it live. Server-render the substance. Speed matters for the same reason, and the work in our note on Core Web Vitals applies unchanged.

Being mentioned elsewhere matters more than it used to

These systems assemble an answer from several sources, so what other sites say about you is part of your visibility, not a side channel. A consistent business name, address and phone number across directories, an accurate map listing, a few genuine third-party mentions, and reviews that are actually on the internet all feed the same picture. For a business serving a specific area, the groundwork in local SEO for Nepali businesses is doing double duty now: it decides both what the map pack shows and what an assistant says when someone asks for a supplier in Kathmandu.

What has not changed

Crawlability, speed, internal links and genuine subject depth still decide whether you are in the candidate pool at all. No amount of answer-shaped formatting rescues a page that loads in six seconds, or a thin page on a topic where you have nothing original to say. The fundamentals were not replaced. A layer was added on top of them, and that layer reads from the same index.

Accept that some traffic is not coming back

Definitional and how-does-it-work queries are the ones being absorbed, and no rewrite recovers them. If a page's entire job was to explain what a domain name is, the assistant now does that and the visit was never worth much anyway. The pages that keep earning clicks are the ones an assistant cannot resolve on its own: your prices, your process, your availability, your work, and anything requiring a decision the reader wants to make with a person. Shift effort there rather than trying to out-explain a model on general knowledge.

Measuring it honestly

Expect impressions to hold or rise while clicks flatten, and expect a slice of your visitors to arrive already informed. That combination looks like failure in a dashboard and often is not. Watch referrals from assistant domains, track branded search volume as a proxy for being named in answers, and check whether the people who do arrive convert at a higher rate than they used to. Fewer, better-qualified visits is a normal outcome, and the enquiry count is the number that settles the argument.

• Impressions and clicks separately, never a single blended traffic figure.
• Referral traffic from assistant domains, as its own segment.
• Branded search volume, month over month.
• Enquiries per hundred visits, which usually rises when informational traffic falls away.
• A monthly manual check: ask the five questions a customer would, in two assistants, and record who gets named.

The practical move this quarter

Take your ten highest-intent pages and rewrite the top of each one so the first passage under every heading answers that heading. Add a dated FAQ block containing the questions your sales conversations actually contain, including the ones about price you have been avoiding. Confirm your robots rules say what you think they say. That is roughly a week of editing on pages you already own, and it is the highest-leverage response available to a small team.

If you would rather have someone work through it with you, that is the ordinary shape of our SEO engagements. And if you are on the other side of this problem, building a product that retrieves and cites documents rather than trying to be cited by one, the same passage-level thinking applies to your own AI and machine learning work.