Search did not disappear when assistants arrived. It moved. A growing share of research questions now end inside ChatGPT, Gemini, Perplexity or a Google AI Overview, and the user reads a synthesised answer with three or four sources attached. Those citations are the new page-one positions, and almost nobody is optimising for them deliberately.
This is a working playbook for generative engine optimisation — how to get your pages quoted and attributed inside AI answers. It is the same process Datalabs runs on client programmes, including what we measure and what we have stopped doing because it did not work.
What actually gets cited
Assistants do not cite pages. They cite passages. A model assembling an answer needs a span of text it can lift more or less intact, attribute to a source, and defend if challenged. That has three practical consequences.
Self-contained beats contextual. A paragraph that only makes sense after reading the two above it cannot be lifted. A paragraph that opens with the answer, names the subject explicitly rather than saying “it” or “this approach”, and completes the thought in two or three sentences can be lifted verbatim. Rewrite for extraction, not for flow alone.
Specific beats comprehensive. “SEO takes time” is unciteable. “Technical SEO fixes typically move rankings within four to six weeks, while content and authority programmes compound over three to nine months” is citeable because it contains a claim a model can attribute to you rather than to general knowledge. Numbers, date ranges, named tools, named platforms and named methods are all citation hooks.
Verifiable beats confident. Models increasingly weight sources that state their basis. “Across 40 client accounts in 2025” or “measured over a 90-day window” gives the passage provenance. Unsupported superlatives — best, leading, world-class — are the fastest route to being ignored.
The content shape that works
Take a question your buyers actually ask, in the words they use with an assistant. Not “enterprise SEO solutions” but “how long does SEO take to work for a new site”. Then structure the page like this:
- H2 phrased as the question, near-verbatim. This gives the retrieval layer an unambiguous match.
- Direct answer in the first sentence under that heading. No throat-clearing, no “it depends”, no restating the question.
- Two supporting sentences with a number, a condition or an exception. This is the part that makes the passage worth quoting rather than merely correct.
- Then the depth — the reasoning, the worked example, the caveats — for the human who keeps reading.
A page built this way serves three audiences at once: the assistant looking for a liftable passage, the classic search engine looking for topical coverage, and the person who arrived and wants the real detail. You do not need separate AI content.
A worked example
Weak, and typical of most agency pages:
Our comprehensive approach to conversion rate optimisation leverages data-driven insights to unlock growth opportunities across your digital funnel.
Nothing in that sentence can be quoted, because it makes no claim. Rewritten for citation:
Conversion rate optimisation needs roughly 1,000 conversions per month across the tested funnel to detect a realistic 5 to 10 percent improvement within a two-to-four-week A/B test. Below that volume, sequential redesigns validated with session recordings and moderated usability testing produce more reliable decisions than underpowered split tests. Most programmes recover their cost in the first two to three months by removing obvious friction before running any experiment.
Three sentences, four extractable facts, one clear position. That passage has been quoted back to us in assistant answers within weeks of publication.
Technical foundations you cannot skip
GEO is not a replacement for technical SEO — it depends on it. Assistants draw from indexes and from live retrieval, and both need crawlable, renderable, fast pages.
- Server-render your content. If the answer only exists after client-side hydration, some retrieval systems will not see it. This single issue disqualifies a surprising number of otherwise strong sites.
- Do not block AI crawlers by accident. Check
robots.txtforGPTBot,PerplexityBot,ClaudeBot,Google-ExtendedandCCBot. Blocking them is a legitimate strategic choice, but make it a decision rather than an inherited default. - Mark up entities.
Organization,Person,Service,Product,ArticleandFAQPageschema help models resolve who you are and what you do. Keep markup identical to visible text; divergence risks a manual action and teaches the model nothing. - Keep facts consistent across pages. If your founding year, service list or headline numbers differ between the homepage, the about page and your LinkedIn profile, the model has no confident value to cite and will quietly prefer a competitor with a coherent record.
Off-site reinforcement
Models weight corroboration. A claim that appears only on your own domain is weaker than the same claim appearing on your site, in an industry publication, in a directory listing and in a conference bio. This is why digital PR has become a GEO tactic rather than only a link-building one — the mention matters even where the link does not.
Priorities, in order of return:
- Original data. A survey, a benchmark study or an anonymised analysis of your own client results gives other publishers something to cite, and citation chains propagate into model training and retrieval.
- Expert commentary. Named quotes from named people at your company, in outlets your buyers read. This builds the entity association between the person, the company and the topic.
- Structured directories and profiles. Unglamorous, but these are high-trust sources for entity resolution. Keep them accurate and identical.
- Community answers. Where your team genuinely participates, substantive answers in forums and Q&A sites are frequently retrieved.
Measuring it
You cannot manage citation share without measuring it, and no platform reports it for you yet. Build the measurement yourself:
- Write a fixed set of 30 to 60 prompts covering your commercial and informational topics, in natural assistant phrasing.
- Run them against each major assistant on a fixed schedule — monthly is enough for most categories, weekly for fast-moving ones.
- Log three things per prompt: is your brand named, is a page of yours cited with a link, and which competitors appear.
- Track the trend, not the absolute. Individual answers vary between runs; the direction over three months is the signal.
Pair that with referral traffic from assistant domains in your analytics, and with branded search volume, which reliably rises when assistant visibility rises even where the click does not happen.
What we stopped doing
Three tactics that sounded plausible and produced nothing measurable:
- Publishing separate “AI-optimised” versions of pages. Duplicate thin variants added crawl waste and no citations. Fix the primary page instead.
- Keyword-stuffing question headings. Assistants match on meaning, not exact strings, so twelve near-identical H2s reads as low quality to both models and human readers.
- Volume for its own sake. Twenty shallow posts a month reduced citation share, because the site’s average passage quality dropped. Depth on fewer topics reversed it.
Where to start this quarter
Pick your ten highest-commercial-intent topics. For each, find the question a buyer would actually type into an assistant, and rewrite the top of the relevant page so the first passage under a question-shaped heading answers it completely, with one concrete number. Add or correct the schema. Then set up your prompt panel and take a baseline before anything else changes.
That is one focused month of work, and it is the highest-leverage content project available to most companies right now — because the surface is new, the competition is thin, and the same changes improve classic rankings at the same time.