What AI search changes about your website

For twenty-five years the deal was stable: you published a page, a search engine ranked it in a list, and a person chose which link to open. Optimisation meant competing for position in that list.

AI assistants change the shape of the interaction. Someone asks a question, gets a synthesised answer, and sees a handful of citations. There is no list of ten to browse. Either your material informed the answer, or it did not.

That sounds like a larger disruption than it is. Most of what makes content usable to an assistant is what makes it usable to a well-informed reader in a hurry. But the emphasis shifts in ways worth being deliberate about.

Answer first, then support it

The conventional structure of a marketing page builds towards its point. Context, then narrative, then the substance somewhere in the middle.

A reader skimming will miss it. A system extracting a factual answer will often miss it too, or lift a partial statement from the preamble.

The alternative is to state the answer in the first paragraph under each heading, then support it. If a section is headed "How long does a Power Platform assessment take," the first sentence should answer that, and the rest should qualify and explain. This is better for human readers as well — it is the structure of a good briefing rather than a good essay.

Say what you do in plain, extractable text

A surprising amount of company information exists only in images, in a video, in a carousel that requires JavaScript to render, or in language too abstract to be extracted at all.

"We deliver transformational outcomes for forward-thinking organisations" tells a reader nothing and gives an assistant nothing to cite. "We build Power Platform applications, Copilot Studio agents, and Power BI reporting for mid-sized organisations" is specific, extractable, and true or false — which is exactly what makes it usable.

The practical test: if someone asked an assistant "what does this company do," could it answer accurately using only the visible text on your site? If the answer depends on inference from imagery or brand language, it will be inferred badly.

Consistency across the web matters more than it used to

Assistants build a picture of an entity from multiple sources — your site, your LinkedIn page, directories, articles that mention you. When those disagree about your name, your focus, or your location, confidence drops and the model hedges or omits you.

This makes basic housekeeping unusually valuable. The same company name, the same description, the same service language across every public profile. It is not sophisticated work, and inconsistency is common enough that fixing it is often a differentiator.

Structured data as clarification, not decoration

JSON-LD does not make an assistant cite you. What it does is remove ambiguity: this string is the organisation name, this is the logo, this is the author, this is the date the article was updated.

Two rules keep it useful. Mark up what is actually on the page — structured data that describes content a visitor cannot see is a spam signal, not an optimisation. And do not add markup speculatively to chase rich results. FAQ markup on a page with no genuine FAQ, or review markup with no reviews, is a liability.

The baseline worth having: Organization on your home and about pages, Article on your insight pages with real author and date information, BreadcrumbList on non-home routes, Service on service pages.

Depth on a topic beats coverage of many

Publishing one article each about twelve subjects produces twelve pages with no authority on anything. Publishing eight connected pieces on one subject, cross-linked and building on each other, produces something that both search engines and assistants can recognise as substantive coverage.

Choose the subjects you want to be known for — realistically, two or three — and go deep. Link related pieces to each other and to the relevant service pages, so the relationship between your expertise and your offering is explicit rather than implied.

The technical foundations did not stop mattering

None of the above helps if the content cannot be retrieved.

• Render content in HTML. If the substance only appears after client-side JavaScript executes, some crawlers will see an empty page. Static or server rendering removes the question.
• Be fast. Crawl budget and user behaviour both respond to page speed, and Core Web Vitals remain a real signal.
• Keep URLs stable and canonical. One canonical host, consistent trailing slashes, and permanent redirects when routes genuinely change.
• Generate a sitemap and keep it accurate. Submit it to Google Search Console and Bing Webmaster Tools; assistants draw on both indexes.
• Use real headings in a logical order. Heading structure is how both assistive technology and extraction systems understand document hierarchy.

Accessibility and AI readability turn out to be closely related problems. Semantic HTML, meaningful headings, descriptive link text, and alternative text serve screen readers and extraction systems for the same underlying reason: they make structure and meaning explicit rather than visual.

What not to do

Two temptations are worth naming.

Hidden text aimed at crawlers. Content served to systems and not to people is the oldest form of search manipulation and is recognised as such. It also fails the basic test of the current environment, where an assistant may quote your hidden text back to a human who then visits and cannot find it.

Generated content at volume without review. Publishing large quantities of unverified generated text is a fast way to accumulate factual errors under your own name. Use AI for research, structure, and repurposing; keep a named human author accountable for what ships. Attribution to a real person with real expertise is increasingly a signal in itself.

The short version

AI search rewards being clear about what you do, being specific enough to be quoted, being consistent about your own facts, being deep on a small number of subjects, and being technically retrievable.

That is not a new discipline. It is the discipline that always worked, with less tolerance than before for vagueness.