AI SEO: A Practitioner's Map of the Generative Engine Optimization Discipline

A direct-to-consumer brand I advised last winter had spent eighteen months and roughly €380k on a content programme — eighty long-form articles, a refreshed information architecture, a fast site, all the classical SEO hygiene. Organic traffic from Google was up 31%. The CEO was happy. The CMO was happy. The procurement officer who renewed the SEO retainer was happy. In the same period, the brand’s citation rate inside ChatGPT for the dozen prompts a competitor analyst had been tracking — “what’s the best [product category] for [use case]”, the prompts the brand’s actual customers were typing into LLMs — went from 4% to 3%. The competitor’s went from 9% to 22%. The brand had won Google’s waning ten-blue-links SERP traffic. They had lost the rising AI-generated answer surface.

That gap is what AI SEO, properly understood, is the answer to. It is not a tactic added on top of classical SEO. It is not a re-skinned content-marketing pitch. It is a discipline that takes seriously the fact that the surfaces shipping AI-generated answers — Google AI Overviews, ChatGPT Search, Claude, Perplexity, You.com, the model layers inside Bing and Copilot — consume content differently from a 10-blue-links SERP, and that the optimisation work for those surfaces is therefore different in shape, different in cadence, and owned by a different part of the organisation. Most published material in this category does not take any of that seriously. Most of it is content marketers writing about AI with the label updated, the techniques unchanged.

This hub is the practitioner-side map. What the discipline actually consists of in 2026, who in the enterprise should own it, what the four real techniques are, what the brand-visibility tooling market looks like, and where the honest claim about most of this is content marketing in a new hat applies and where it does not. The deep-dives underneath cover the operational questions: how to rank in ChatGPT specifically, how to rank in Google AI Overviews specifically, the brand-visibility tooling comparison, and the structured-data work that anchors all of it.

What is actually different about AI search surfaces

Start with the mechanism. A classical Google SERP returns ten links; the user picks one or more; the user does the synthesis. An AI search surface — whether it is Google’s AI Overviews, ChatGPT Search, Perplexity, or Claude with web browsing enabled — fetches a candidate set of sources, reads them through the model, synthesises an answer, and emits the answer with citations. The user reads the answer. They click a citation roughly 12-18% of the time on average across surfaces, with substantial variance by query type — informational queries cite-through at the low end of that range, commercial queries closer to the higher end. The click-through is no longer the load-bearing metric; the citation itself is.

This changes three things about the optimisation work, and the changes compound.

One. The candidate set is narrower. Where a SERP shows ten organic results, an AI Overview typically grounds against three to seven sources, ChatGPT Search against a similar range, Perplexity slightly wider. The ranking competition is therefore not “top ten or invisible” but “top five or invisible,” with the cliff between position five and position six steeper than it has ever been for organic Google traffic. The implication for site-level strategy is direct: depth in fewer topics beats breadth across many, because depth is what gets you into the top-five candidate set on a query that matters.

Two. The model is reading you, not the user. This sounds obvious; the operational consequences are not. Content that ranks for humans by being engaging — narrative openers, emotive framing, brand-voice flourishes — does not necessarily get cited by a model, which is parsing for declarative claims and structured assertions it can quote. Conversely, content that is dense with structured data, comparison tables, named-source citations, and Q&A blocks — the unglamorous structured-content shape — outperforms in AI citation rate in a way that often inverts its classical-SEO ranking. I have seen brand pages with mediocre Google rankings get cited consistently in ChatGPT because their FAQPage schema and comparison tables were the cleanest in the candidate set.

Three. Citation is entity-anchored, not URL-anchored. When a model decides to cite, it cites because the source-page entity matches the query entity it has been asked about. The entity-level work — making sure the model has learned that your brand is a credible source on this specific topic across the training data, the tool-call surfaces, and the linked Web — matters more than any single page’s optimisation. This is the part of the discipline that looks least like classical SEO and most like reputation management. It is also the part where the published advice is thinnest.

The four techniques that actually move the citation rate

If you have read enough of the published material on GEO or AI SEO to be cynical, you have noticed that most of the advice collapses into “write good content with structured data.” That is not wrong; it is just inadequately specific. The four techniques below are the ones I have watched move the citation rate measurably, in engagements where the team baseline-measured before and after.

Technique one: structured data and LLM-citable content shapes. The boring one, and the one with the highest leverage in 2026. The schemas that earn AI citation are not exotic. FAQPage for question-answer content. Article with full author, datePublished, dateModified, image, and headline properties. HowTo with discrete HowToStep blocks for procedural content. BreadcrumbList for navigation context. Product with Offer for commerce pages. Comparison (via custom WebPage mainEntity with structured comparison tables) for “X vs Y” content. The honest part: most enterprise sites in 2026 have between 30% and 60% of the schema coverage they should have, with the gap concentrated in the shapes that AI surfaces specifically quote. Google’s own structured data documentation remains the cleanest reference, and the AI Overviews team has been explicit in their guidance that schema is a signal they weight.

The content-shape side is equally specific. Short paragraphs (under 80 words) with declarative claims. Direct answers in the first 50 words of any section that targets a “what is X” query. Comparison tables with consistent column structure across rows. Glossary-style definitions where the definition is one paragraph, not three. These shapes are not aesthetic preferences. They are the shapes the model parses cleanly and quotes verbatim, and pages that adopt them get cited at materially higher rates than pages that hide the same claims inside narrative prose.

Technique two: source authority and E-E-A-T signals. The same E-E-A-T work that survives traditional SERPs is what survives AI surfaces, and it survives more aggressively. Named human author, not “Editorial Team.” Author bio that establishes domain expertise — real credentials, real experience, not LinkedIn-bio platitudes. Published and updated dates that are real, not build timestamps. External citations to authoritative sources backing factual claims. First-person experience markers where applicable. The model treats these as trust signals because the training and reranking pipelines treat them as trust signals; the alignment between what classical search rewards and what AI surfaces reward is high on E-E-A-T specifically, and the teams that have invested in the byline-and-bio work over the last two years are now compounding that investment in citation share.

The connection to the governance hub is direct. The same evidence-trail discipline that AI governance demands — named owner, dated artefact, traceable provenance — is what AI search surfaces reward when they decide which source to cite. The brand that has built governance hygiene into its content programme as a side effect of compliance work has accidentally built the most-cited content programme on its competitive set. I have seen this happen twice. Both times the brand did not notice for six months because they were not measuring the right surface.

Technique three: cross-publication on the surfaces LLMs index. This is the technique that distinguishes practitioners from content marketers. LLMs do not just read your website. They read Reddit, Hacker News, GitHub Discussions, Substack, named industry blogs, Wikipedia, the major Stack Exchange properties, the major review sites in your category, and a long tail of forum and community surfaces. The entity associations the model builds for your brand are built from the cross-surface mention pattern, not from any single property. A brand that publishes only on its own domain is invisible to the model on every prompt the model resolves by checking what other sources say about that brand — which is most prompts.

The operational implication is that an AI SEO programme has to include a deliberate cross-publication strategy. Engagement on relevant subreddits, technical content on GitHub for software-adjacent brands, named-author guest posts on industry properties the model treats as authoritative, contributions to Stack Exchange where the brand has subject-matter expertise, and a real Wikipedia presence where the notability bar is met. None of this is paid placement. All of it is the unglamorous work of being a genuine participant in the surfaces your customers and the models you want to cite you are already reading.

Technique four: prompt-anchoring and entity association. The most-overlooked technique, and the one closest to the model’s actual reasoning. The specific phrasings users say to LLMs become the entity associations the model builds. If your customers ask LLMs “what’s the best AI strategy framework for a mid-sized European bank,” and your brand is mentioned in proximity to that exact phrasing — in your own content, in cross-publications, in the press, in named analyst commentary — the model learns the association. The prompt-anchoring work is the deliberate identification of the prompts that matter for your business, the audit of how often your brand currently surfaces against them, and the structured campaign to increase that surface rate through technique-three cross-publication and technique-one content shaping.

This is the work the brand-visibility tools comparison page addresses directly. Tools like Profound, Otterly, Peec AI, and Goodie AI sample the major LLM surfaces against a brand’s prompt set on a recurring cadence and produce a citation-rate baseline. The tools are useful for measurement; the optimisation work that responds to the measurement is the four techniques above, in that order of leverage.

Who in the enterprise owns this work

The org-design question is where most AI SEO programmes in 2026 are mis-shaped. The default answer — “marketing owns SEO, so marketing owns AI SEO” — produces predictably weak outcomes, because the work is not predominantly marketing work. The four techniques above are, in order: an engineering task (structured data), an editorial-and-credentialing task (E-E-A-T), a community-and-PR task (cross-publication), and a strategy-and-analytics task (prompt anchoring). Marketing is one of the four owners, not the load-bearing one.

The pattern I have seen work, across three engagements where the brand measurably moved their AI citation rate inside twelve months, is this. A named technical SEO lead, reporting into the engineering or platform organisation, owns the structured-data work, the schema audit, the evaluation harness against brand-visibility tooling, and the cross-surface technical implementation. The CMO’s office owns the editorial-calendar input and the brand-voice consistency across cross-publications. A communications or PR lead owns the named-publication relationships and the Wikipedia work where applicable. The CIO or CTO is the accountable executive at portfolio level, because the budget for the engineering work sits there.

The pattern that fails, predictably, is the one where the CMO buys a brand-visibility tool, hires an agency to run it, and treats the structured-data work as a website-team backlog item that gets prioritised behind every other request. The agency reports a citation-rate number every month. The number does not move because the underlying technical and cross-publication work does not happen. The contract renews twice before someone notices. I have audited two of these engagements and walked away from a third because the org-design fix required a conversation with the CMO the CEO was not ready to have.

The connection to the capabilities cluster is the same one made for AI-SRE and for orchestration work. The capability layer is engineering work that requires marketing input, not marketing work with engineering as a service desk. Get the reporting line right or the citation rate will not move.

The brand-visibility tooling market

The market for tools that measure LLM citation rate is roughly two years old, growing fast, and not yet consolidated. The major names in mid-2026 — Profound, Otterly, Peec AI, Goodie AI, the AthenaHQ-adjacent work, the early Ahrefs and Semrush features moving into the space — all do approximately the same job at the core: sample a defined prompt set against a defined model surface set, parse the responses for brand mentions and citations, report the rate over time. The differences are at the edges: prompt-set size, surface coverage (some are ChatGPT-only, some cover Perplexity and Claude, the broader ones cover AI Overviews where Google permits the sampling), reporting cadence, integration with the rest of the marketing stack.

The honest procurement read, which the brand-visibility tools comparison page covers in operational depth: buy the cheapest defensible thing on a per-prompt or per-seat tier, use it as a measurement baseline for three to six months, and decide whether the citation-rate signal is moving the procurement conversation forward. If yes, expand on the same vendor. If not, churn — most of these tools are priced for churn-tolerance specifically, because the market is consolidating and the vendors know it. Signing a multi-year contract in this category in 2026 is the same procurement mistake as signing a multi-year AI-SRE contract: the category is moving too fast for the term length to make sense.

What the tools do not do, and what no tool will do for you, is the optimisation work itself. They measure. The four techniques above are what moves the measurement. Buying the tool and skipping the techniques is the most expensive variant of the same mistake the failed engagements above made — a measurement layer with nothing behind it.

Where the honest claim about content-marketer rebrand applies

Most published advice on AI SEO or GEO in 2026 is content marketing with a new label. The tell is the technique stack: “write helpful, comprehensive content,” “use structured data,” “build topical authority,” “earn backlinks.” All of these are correct. None of them are different from the SEO advice published in 2018. The discipline-defining technique, the one that actually distinguishes AI SEO from classical SEO, is the cross-publication and prompt-anchoring work above — technique three and technique four. Most content-marketer-authored pieces leave both out, because both fall outside the content-marketing remit.

The discipline is real where the advice goes specific. Anthropic has published guidance on how Claude resolves source citations that is dry, technical, and exactly the kind of source-grounding work that maps to technique one. OpenAI’s published material on how ChatGPT Search ranks and cites sources, while less detailed than the academic GEO literature, is similarly technical when it gets specific. The 2023 Princeton/Georgia Tech Generative Engine Optimization paper — the academic paper that gave the discipline its name — is still the cleanest technical reference for what actually moves citation rate, and it is more rigorous than 95% of the published industry material that came after it. Search Engine Land’s ongoing AI SEO coverage is the strongest industry source, with explicit attention to the surface-specific differences between AI Overviews, ChatGPT Search, and Perplexity that the broader trade press tends to flatten.

The discipline is not real where the advice stays generic. The brochure-shaped output — Top 7 Ways AI Is Transforming SEO, the ungated PDF with the email gate, the agency that promises to “make your brand AI-ready” without naming a single structured-data change — is content marketing with the AI SEO label, sold to a market that is anxious enough to buy almost anything wearing that label in 2026. The same procurement-anxiety dynamic that the enterprise AI literacy piece names for the EU AI Act remediation market is operating here. The fix is the same: name what you are actually buying, evaluate against the four techniques above, and refuse to renew anything that is not measurably moving the citation rate.

How this hub connects to the rest of the site

Read the parent capabilities hub for the broader frame of what capability work means versus strategy work; AI SEO sits firmly on the capability side, even though it has strategy implications for any brand whose discovery surface includes LLMs. Read the enterprise AI literacy article for why the procurement and middle-manager tier matters specifically — most AI SEO procurement decisions get made by the tier-three population the literacy programme has to reach. Read the governance hub for the E-E-A-T evidence-trail discipline that overlaps directly with what AI search surfaces reward.

The deep-dives underneath this hub address the operational questions one surface at a time. How to rank in ChatGPT is the ChatGPT-Search-specific play, including the Bing-index dependency that most coverage misses. How to rank in Google AI Overviews is the AI-Overviews-specific play, including the organic-ranking dependency that determines which sites are even in the candidate set. The forthcoming brand-visibility tooling comparison and the structured-data deep-dive cover the measurement layer and the technical implementation respectively.

If you are starting this work from a blank page, the order I would run it in is: baseline-measure your current citation rate with a small-tier brand-visibility tool for six weeks; audit your structured data coverage and fix the gaps; rewrite the three to five pages that target your highest-value commercial queries into LLM-citable content shapes; design the cross-publication programme for the surfaces relevant to your category; and only then start the prompt-anchoring campaign at scale. Most teams want to start with the prompt-anchoring work because it sounds the most like marketing. That order produces a programme that looks busy and does not move the citation rate. The order above produces a programme that moves slowly for the first quarter and then compounds.

---

Sources & methodology

• Generative Engine Optimization (GEO), Aggarwal et al., Princeton/Georgia Tech, November 2023 — the founding academic paper on optimising for generative engines; still the cleanest technical reference
• Google — Structured data introduction — schema reference for the structured-data technique
• Anthropic — Claude 3 model family — source-grounding and citation behaviour reference
• Search Engine Land — AI SEO coverage — strongest industry source, surface-by-surface
• Methodology: citation-rate baselines and movement data drawn from fractional CTO advisory engagements (2024–2026) with brand-visibility tooling (Profound, Otterly, Peec AI sampling) on defined prompt sets per engagement; engagements anonymised by sector and headcount, prompt sets published per-engagement under NDA

The four-technique scoring rubric is CC-BY-4.0 and will live on the brand-visibility tooling page once it ships. If a cited claim looks wrong, send it and I will publish the correction with attribution.