What should I optimize to rank in both search engines and AI assistants?
This is the page-level answer rather than the strategic one. For each element on a page, what it does for rankings, what it does for citations, and what to write.
This is the page-level answer. For each element on a page: what it does for rankings, what it does for citations, and what to actually write.
Element by element
| Element | Search | AI assistants |
|---|---|---|
| Server-rendered content | Optional (Googlebot renders) | Required |
| Title tag | High — drives click-through | Low |
| Meta description | Moderate — click-through only | Negligible |
| Question-led headings | High | High |
| Answer in first two sentences | High — snippets | Very high |
| Self-contained passages | Moderate | Very high |
| Specific dated claims | Moderate | Very high |
| Structured data | Moderate | Moderate |
| Internal links | High | Moderate — aids retrieval |
| Page speed | Low — tiebreaker | Moderate — crawler timeouts |
Rendering
The only element where the two channels genuinely disagree, and the disagreement is absolute.
Googlebot renders JavaScript in a headless Chrome engine, so client-side content ranks fine. AI crawlers do not — Vercel and MERJ recorded zero JavaScript execution across more than 500 million GPTBot fetches, with GPTBot downloading JavaScript files roughly 11.5% of the time and never running them. ClaudeBot and PerplexityBot behave the same way.
What to do: server-render anything that constitutes an answer. Load your page with JavaScript disabled and read what remains. That is your entire AI-visible surface.
Title and description
Search: the title is the highest-leverage single field on the page, because it drives both relevance understanding and click-through. Use the searcher’s phrasing, front-loaded, under about sixty characters. The meta description is not a ranking factor and is a click-through factor — write it as a promise rather than a summary.
AI: minimal effect. There is no snippet to click, so the persuasive function disappears entirely. The title still signals topic, which helps retrieval marginally.
Conclusion: keep doing this properly for search. Do not expect it to move citations, and do not skip it because it does not.
Headings
Both channels, strongly. Phrase headings as the questions people actually ask. Retrieval matches against them and readers scan them.
One H1 matching the page’s subject. H2s as real sub-questions. H3s only where a sub-question genuinely subdivides. Do not skip levels for visual reasons — it is a minor accessibility issue and a minor parsing one.
The test for a good H2: could someone type it into an assistant verbatim? “How much does keyword research cost?” passes. “Cost considerations” does not.
Check a page against both channels
Klepha’s audit checks retrievability and extraction quality, and shows which pages assistants cite in your category.
Run my free scanThe first paragraph
The highest-value real estate on the page for both channels.
Two to three sentences containing the direct answer to the question the heading asks. No preamble, no history, no throat-clearing. Context goes after, not before.
For search this is what a featured snippet frequently takes. For AI it is the passage most likely to be extracted. For a human it is the difference between staying and returning to the results page. One edit, three benefits — which is why this is the single most valuable change available on most pages.
The body
Self-contained paragraphs. Each should survive being copied out of context. Name subjects rather than referring back with pronouns. This reads marginally more repetitive to someone consuming the whole page and dramatically better to everyone else.
Specific, dated, attributable claims. The strongest single editorial lever for citation. The Princeton and IIT Delhi GEO study found Statistics Addition, Quotation Addition and Cite Sources were the three highest-performing methods tested, at roughly 30–40 percent relative visibility improvement across around 10,000 queries. Name sources, date figures, and include at least one number nobody else has.
One topic per page. A page answering six questions answers none cleanly enough to be preferred. This helps citation strongly and rankings moderately, since it also reduces cannibalisation.
Tables and lists where the content genuinely is structured. Comparisons, steps and criteria belong in these formats and are disproportionately likely to be extracted.
Structured data
Article, FAQPage, HowTo and Organization where genuinely applicable. Moderate value in both channels, and one absolute rule: it must mirror the visible text exactly.
Markup contradicting the page is worse than no markup, because it signals either carelessness or manipulation. And schema on a page no crawler can read does nothing at all, which is why this belongs after rendering rather than before it.
Links
Internal links with descriptive anchor text. High value for search — they distribute authority and establish topical relationships. Moderate for AI, where they aid retrieval by making pages discoverable.
Outbound links to authoritative sources are the counterintuitive one. They cost nothing in rankings and help citation meaningfully, because a page that names its sources reads as analysis rather than assertion, and a model quoting you inherits your citations as supporting evidence.
New pages versus existing ones
Most of this work belongs on pages that already exist, and the reason is arithmetic rather than preference.
An existing page ranking at twelve has proven relevance, accumulated internal links, some age, and possibly external links. Restructuring it answer-first and adding a concrete claim costs an hour and operates on demand that already exists. A new page starts from zero on every one of those dimensions and needs months before it can be judged.
The rule of thumb that holds up: restructure before you write. Only publish new pages against questions nothing on your site currently addresses — and verify that by searching your own domain for the query rather than assuming, because most sites have more coverage than the people working on them remember.
Six element-level mistakes
Writing the title for the brand rather than the query. “Acme | Solutions for Modern Teams” tells a searcher nothing and matches nothing. Front-load what the page answers.
Opening with context. The single most common structural failure. Three paragraphs of background before the answer costs you the snippet, the citation and the reader simultaneously.
Using headings as labels rather than questions. “Overview”, “Considerations”, “Next steps” match nothing anyone types or asks.
Building comparison content as prose. If the content genuinely is a comparison, a table is both easier to read and far more likely to be extracted. Prose paragraphs describing four options against three criteria serve nobody.
Adding schema that does not match. FAQPage markup with questions that do not appear on the page is a trust liability rather than a shortcut, and it is easy to detect.
Assuming the CMS renders server-side. Many modern stacks do by default and some do not, and component-level client rendering can hide inside an otherwise server-rendered page. Test the actual output rather than trusting the framework’s reputation.
What to do about pages that cannot serve both
Some pages genuinely cannot. A pricing page has to persuade, and persuasion and extractability pull in different directions. A product landing page optimised for conversion should not be restructured into a question-and-answer document.
The resolution is not to compromise the page. It is to accept that commercial pages are conversion assets rather than citation assets, and to build the citation surface separately — documentation, guides, comparison pages and data. A company whose entire site is landing pages has no citation surface at all, and the fix is adding one rather than degrading what already works.
This also clarifies a common confusion. When an assistant does not cite your homepage, that is expected rather than a failure. Homepages are almost never cited, because there is no extractable factual claim in them.
The edit order for an existing page
When improving a page that already exists, the order matters because each step makes the next easier to judge. Confirm it renders server-side. Move the answer up. Split it if it covers several questions. Fix the pronouns so passages stand alone. Add one concrete claim. Then, last, adjust the title and schema — because those are quick and there is no point polishing them on a page that is about to be split in two.
A page template that works for both
- H1 matching the question, in the searcher’s phrasing.
- Two to three sentences answering it directly. No preamble.
- A short takeaways box — five bullets, each self-contained and specific.
- H2 sub-questions, each followed immediately by its two-sentence answer, then detail.
- At least one table or numbered list where the content genuinely is structured.
- At least one dated, sourced statistic and ideally one number that is yours alone.
- An FAQ block — questions as headings, two to three sentence answers, matching FAQPage schema.
- Internal links to related pages with descriptive anchors.
- A visible last-updated date, honest.
- All of it server-rendered.
One note on applying it: the template describes an informational page. Commercial pages should not be forced into this shape, because persuasion and extraction pull in different directions and a pricing page rewritten as a question-and-answer document converts worse without becoming meaningfully more citable.
That structure serves a skimming reader, a featured snippet, and an AI extraction pipeline simultaneously — which is possible because all three want the same thing: the answer, early, stated plainly, specific enough to be useful.
Sources
- Vercel and MERJ — analysis of 500M+ GPTBot fetches finding zero JavaScript execution; GPTBot downloaded JS files ~11.5% of the time without running them; same behaviour for ClaudeBot and PerplexityBot.
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande — “GEO: Generative Engine Optimization”, ACM SIGKDD 2024 (arXiv:2311.09735); ~30–40% relative visibility improvement across ~10,000 queries.
- Google Search Central — structured data guidelines and rendering documentation.
Frequently asked questions
What should I optimize to rank in both search engines and AI assistants?
Rendering first, since AI crawlers cannot read JavaScript-only content and Googlebot can. Then question-led headings, an answer in the first two sentences, self-contained body passages with specific dated claims, structured data matching the visible text, and descriptive internal links. Titles and meta descriptions matter for search click-through only, and are still worth doing.
Do title tags matter for AI search?
Much less than for search. Assistants do not present a clickable snippet, so a title’s persuasive job disappears. It still helps engines understand the page topic, so write it for search click-through and accept that it does little for citation.
Should headings be questions?
Yes for both channels. Question-led headings match how people search and how they prompt, and they create clean question-and-answer blocks that both featured snippets and AI extraction can lift. “How do AI engines choose sources?” works; “Citation mechanics” does not.
How long should the first paragraph be?
Two to three sentences containing the direct answer. This is the highest-value real estate on the page for both channels — it is what a skimming reader sees, what a featured snippet often takes, and what an AI engine is most likely to extract. Context and background belong after it, not before.
Does schema markup help AI assistants?
As a supporting signal. It gives engines labelled, unambiguous facts and improves confidence in attribution. The requirement is that it mirrors the visible text exactly — markup that contradicts the page is worse than none, and markup on a page no crawler can read does nothing at all.
What single change helps both most?
Moving the answer to the first two sentences of each section. It improves featured snippet capture, helps skimming readers, and is the single strongest determinant of whether an AI engine can extract a usable passage. One edit, three benefits.