GEO · Generative Engine Optimisation
AI search has four ways to miss your brand. We engineer every one.
Right now a buyer is asking AI which vendor to shortlist. The answer has almost nothing to do with the ranking your team checks every week.
Every AI answer goes through three stages before your brand ever gets mentioned, and you can lose at any one of them.
The grounding pipeline · where brands drop out
Explore → Verify → Cite
Explore
Verify
Cite
- 01 Training data gaps Fails at: In the training corpus
- 02 Crawler blocking Fails at: Crawlable by AI agents
- 03 Hallucination Fails at: Represented accurately
- 04 Citation absence Fails at: Cited in the answer
"Which agency should we use for AI search visibility?"
Based on available sources, the most frequently referenced options are Competitor A1, Competitor B2 and Competitor C3, which appear consistently across industry roundups and comparison guides.
Sources it drew from
You rank #1 on Google for this query. It made no difference to the answer.
01. The argument
Ranking on Google is not existing in AI search.
Google's model is Crawl, Index, Rank. AI search runs Explore, Verify, Cite. A page that ranks is not automatically a page that gets cited. Two different jobs, two different builds.
AI search
Explore Verify CiteRoughly 12% of what Google shows overlaps with what ChatGPT actually cites for the same query. About 7% of AI citations go to pages that never made the Google top 10 for that query. The grounding pipeline does not follow ranking order. It follows how connected an entity is inside the model's own map of trusted sources.
Most brands are investing in one channel while their buyers form the shortlist on the other.
Retrieval and ranking produce candidates, not answers. The grounding pipeline adds six stages on top: evidence selection, evidence construction with attribution, constrained generation, cross check and verification (Microsoft Critique), multi model judging (Microsoft Model Council), then answer generation with citation. Critique and Model Council are adversarial review layers that your content has to survive before it is cited. That is why "good content" and "citable content" are not the same document architecture.
Sources
Volpini synthesis, SEO Week. Madhavan synthesis, Microsoft SEO Week.
02. The four failure modes
AI search does not have one failure mode. It has four.
You are not in the model's memory.
Common Crawl releases data at scale every month and is foundational training data for GPT, LLaMA and others. It strips schema markup on ingestion. If your content is not in there, the model has nothing to recall about you.
Not invisibility. Absence.
The only way to find this gap is to probe the model directly with live API access. No crawl dashboard on earth can tell you what a model remembers and what it does not.
Where the brand drops out
Fails at: In the training corpus
AI cannot index what it cannot reach.
GPTBot, ClaudeBot, Google-Extended, PerplexityBot and Bingbot all honour robots.txt and meta directives. One block removes you from that system's grounding pool entirely. Most of these blocks went up in 2023, a quick defensive move while nobody was sure what AI crawlers even were, and most teams never went back to check.
A line in robots.txt is deciding your AI visibility.
Freshness is no longer a ranking boost, it is an eligibility gate, closer to pass or fail than to a bonus. Stale pages get filtered out before grounding ever considers them. A real refresh cadence plus IndexNow keeps evergreen pages eligible. Changing the year in the footer from 2025 to 2026 does not count.
Where the brand drops out
Fails at: Crawlable by AI agents
AI misrepresents your brand at the decision point.
A fact sitting on your page does not mean the model recalls the right version of it. In one live audit, the house edge on a client's blackjack page was 0.57%. Probed without live page access, ChatGPT answered 0.50%. Wrong number, stated with total confidence, read by a buyer mid evaluation.
Worse than absence: a confident wrong answer.
Prose with no entity structure pushes the model back to brute force vector similarity. Schema markup alone barely moves it. In WordLift's benchmark on Vertex AI Vector Search 2.0, navigable entity pages lifted retrieval accuracy 29.6% over schema markup alone (arXiv:2603.10700).
Where the brand drops out
Fails at: Represented accurately
You rank on Google. AI still does not recommend you.
Ranking authority and citation authority run on different infrastructure. You can sit at position one on Google and still be absent from every AI answer, because the model picks sources by how well connected they are inside its own trusted network, not by ranking order.
Two retrieval systems. Most brands engineer one.
Closing that second gap is what the AI Citations service does, placing your brand inside the co-citation clusters the model already trusts.
Where the brand drops out
Fails at: Cited in the answer
03. What we audit
We audit whether AI can reach, read and trust your site.
Four gates, and the grounding pipeline runs them in a strict order. Fail one and you are disqualified before the next one runs. Every gap we find gets a priority and a fix. The order is not arbitrary: the gate that blocks the most closes first.
| Failure mode | Gate it fails | What we fix |
|---|---|---|
| Training data gaps | 01 Discoverable | Corpus presence, entity recognition across foundation model training data, IndexNow on publish, meaningful update and deletion |
| Crawler blocking | 01 Discoverable | robots.txt and meta directives for the five AI crawlers, clean canonical chain, valid sitemap, optional llms.txt declaring your canonical entities |
| Hallucination | 02 Parsable03 Groundable | Server rendered JSON-LD on every meaningful page, self contained passages, clear provenance, no contradictions on the same page |
| Citation absence | 04 Trustworthy | One permanent URI per entity used identically site wide, named authors with external verification, multi surface presence, fact consistency over time |
Gate 01 · Discoverable
The crawler has to reach you.
Clean URLs, valid sitemap, IndexNow, no accidental blocks, clean canonicals. Most enterprise sites carry at least one block nobody still working there knows about. 50% of clicked newly indexed Bing URLs come from IndexNow.
Madhavan synthesis, Microsoft SEO Week
Gate 02 · Parsable
The model has to read what it finds.
Semantic HTML, valid JSON-LD present in the raw server rendered source and not just in what a browser shows you, alt text, video transcripts, clean heading hierarchy, content that does not need JavaScript to exist.
Gate 03 · Groundable
Your content has to work as a citable source.
Passages that answer one question without borrowing context from another page, provenance on every claim, no page contradicting itself. This is the gate most well ranked sites fail.
Gate 04 · Trustworthy
The model has to verify your authority.
The same permanent URI for an entity on every page, character for character, plus named authors, external corroboration and facts that hold steady over time.
Standard audit tools read rendered text and cannot see raw <script type="application/ld+json"> tags. "No schema found" from those tools is a systematic false negative, not a measurement. We verify structured data at source every time, with methods that actually parse JSON-LD. Get this wrong and every structured data investment you have made looks like zero on the report.
04. The entity layer
Schema markup alone does not make you citable.
Every entity on your site needs one permanent URI, and that exact URI has to appear every time you mention the entity, on every page. Name the same thing three different ways across the blog, the product pages and your schema, and the model can no longer tell they are the same entity. Three weak signals scattered around instead of one strong one (Volpini, SEO Week).
The highest leverage change
The highest leverage single on page change is @id consistency. It is nearly free and it shows up in measurement immediately.
Step 01
Identify the 50 to 500 things your site is actually about: products, people, places, concepts. Output is a flat list grouped by type.
Step 02
Assign a permanent URI pattern (/entity/[slug]) and hold to it. An @id is a primary key. Changing it is a migration.
Step 03
Build one entity page per core entity: @type, stable @id, name, alternateName, description, sameAs links to external authorities, subjectOf back edges to every content page about it.
Step 04
Resolve Wikidata Q numbers for every globally notable entity and add them to sameAs. Proprietary entities that are not on Wikidata stay as local URIs.
Step 05
Make @id consistent site wide, and keep it that way with a CI test that flags an inconsistent @id on every pull request.
An entity page tells anyone, human or machine, what a thing is. A content page tells a user how to use it. Hub and spoke, the pattern Wikipedia uses. They target different intents and they do not cannibalise each other. Skip the entity layer and the model has labels with no edges to traverse.
05. How we measure it
We measure what AI says about you, not what a dashboard estimates.
- We run 45 live prompts about your brand through the ChatGPT API with live web search on, so we see the answer a buyer sees: category evaluation, brand comparison, feature specific, use case and objection handling queries.
- One appearance is not a citation. It counts when it appears in three consecutive independent calls.
- Every response is scored against ground truth taken from your own pages.
Fix order is set by commercial query volume and gap severity.
Direct API access, not scraping, so the results match what buyers actually see. Localisation variance is removed by running through a fixed regional proxy. Where a specific fragment on a page is feeding a wrong answer, fragment level data-nosnippet directives exclude that fragment from extraction while the page stays eligible for grounding. Precision control, not a blunt noindex.
06. The research behind it
Our authority is built on data, not LinkedIn.
AI search is a data science problem. Most agencies are not data science companies. We are.
AI citations analysed, the deepest study of its kind
of AI citation research behind the methodology
research partners: DataForSEO, SurferSEO, SearchAtlas
proprietary tools, built for GEO, not adapted from SEO
Three cosine similarity signals measured per citation: site level topical alignment, semantic similarity of the domain to the query, and page level semantic match. Spearman correlation against Harmonic Centrality to map citation frequency against network position. Ordinal regression on citation rank, and cited versus not cited against SERP controlled negatives. A follow on study is running with SearchAtlas and SurferSEO.
AI does not cite the loudest brand. It cites the most connected one.
A small, recurring set of sources gets reused across nearly every answer in a category. Most citation programmes try to appear in every answer. The goal is to become one of the recurring set.
07. The case
Named everywhere. Cited nowhere.
We looked at the world's largest crypto sportsbook. AI assistants named the brand in 89% of responses. Its own site was cited in 11% of them. Affiliates and review sites controlled the rest.
Brand mention rate in AI search
How often AI cited the brand's own website
Own domain citations on category queries ("best crypto sportsbook") at baseline
Everyone knew the brand. AI did not trust the brand's own site enough to cite it. The fix is not more content. It is making your own site the source AI trusts.
What the audit found
Our first report said the site had no structured data. We retracted that: the tools were reading rendered text and could not see the raw tags. The real picture was a sophisticated entity graph broken by two bugs. On most templates the site level node was double encoded as a string inside a JSON-LD array, invisible to every parser, taking the sponsor roster, the Wikipedia link and the author references down with it. And every entity used a bare fragment ID, so the same entity got a different absolute @id on every page. One serialization fix resurrects the graph. Minting absolute, reused @ids consolidates it.
Entity reconciliation · 217 entities by declaration status
- 124 declared valid
- 22 declared bare
- 21 declared broken
- 31 mentioned only
- 19 undeclared
Not one of the 217 is linked to Wikidata, even though 89 of the Q numbers are already resolved and ready to drop into sameAs.
| Gate | Verdict |
|---|---|
| 01 Discoverable | Amber |
| 02 Parsable | RedSchema present but broken, visible only in view source |
| 03 Groundable | RedHub pages near empty of extractable text |
| 04 Trustworthy | RedNo Wikidata grounding, and the model volunteers regulatory caveats on brand legitimacy queries |
The entity work is what should move the 11%. We re-baseline at Week 13.
08. The engagement
Four phases. One complete picture of your AI visibility.
Phase 01
Foundation and benchmarking
What AI knows about you right now.
Phase 02
Content and knowledge graph
What AI understands about your authority.
Phase 03
Model probing and hallucination detection
What AI says about you, versus what is true.
Phase 04
Citation engineering
Who AI recommends, and how to get you in the room.
A 3 month project with a defined outcome, not an open ended engagement that never quite finishes. We re-baseline at Week 13 against the same prompt database, so the change is measured the way the baseline was.
Scope
GEO owns the on page layer: crawler compliance, corpus eligibility, entity schema, @id consistency and hallucination correction. Off page placement, the co-citation clusters and the authority network programme, belongs to AI Citations. Phase 4 identifies exactly where the gap is and hands over the outreach target list. On page work creates the eligibility. Off page work does the placement.
09. What you receive
Eleven deliverables. Your citation probability, and exactly how to engineer it.
| # | Deliverable | What it tells you |
|---|---|---|
| 1 | Target Prompt Database | 100 prompts across category evaluation, brand comparison, feature specific, use case, objection handling and competitive displacement. Drives the baseline and every re-baseline. |
| 2 | AI Crawler Compliance Audit | Every block keeping you out of a grounding pool, prioritised for removal, plus llms.txt and AI info page assessment. |
| 3 | Industry Corpus Inclusion Scorecard | Whether the models have you at all, scored by entity type: brand, product, people, concepts, categories. |
| 4 | Training Data Coverage Report | Per entity recall across the 100 prompt database: what the model knows, what it approximates, what it has never seen. |
| 5 | Training Data Gap Map | Entities where model recall is absent or wrong against your own pages, ordered by commercial query volume. |
| 6 | AI Visibility Monitoring Dashboard | Your own domain citation rate over time against competitors, built on direct API access rather than scraping. |
| 7 | Competitive Knowledge Graph and Topic Gap Analysis | The entities your competitors own pages for and you do not, and which of those gaps cost you citations. |
| 8 | Query Fan-Out Prompt Database | Your top commercial queries expanded across five grounding lanes: aesthetic, popularity, engagement, reviews, freshness. Cover one lane and you miss four paths. |
| 9 | Hallucination Risk Assessment | The 45 prompt benchmark scored against your ground truth, with a fix priority list per entity. |
| 10 | Citation Cluster and Centrality Map | How your brand connects, or fails to connect, to the sources AI treats as authoritative in your category. |
| 11 | Co-Citation Gap Report and Outreach Target List | Which high centrality sources co-cite your competitors and not you. This is the handoff list to AI Citations. |
Behind deliverable 10
The outreach target list is not a guess. It is priced. 248,000+ URLs in the database, each carrying a real acquisition cost against its citation cluster value. That is what lets us tell you which gaps are worth closing and which would cost more than they could ever return. You get the prioritised list, not the raw chart.
10. Closing
Every brand we audit finds a gap. The only question is whether you find it before your buyers do.
We will show you exactly what AI is saying about your brand, where it is getting you wrong, and what to fix first to move your citation rate.
The brands building AI visibility now will own it. The ones waiting will buy it back at a premium. Brie Moreau, Founder, White Light Digital Marketing
We have never run one of these audits and come back empty-handed. If yours is the first, we will say so.