If you are not tracking AI visibility, you do not know there is a problem. But your buyers do.
AI search and Google search are different systems. Ranking on Google does not mean you exist in AI search.
Most brands are investing in one channel while buyers form shortlists on the other.
Why brands that rank on Google are invisible in AI search
Google’s model: Crawl, Index, Rank. AI search: Explore, Verify, Cite. A page that ranks is not necessarily a page that gets cited. These are two different jobs.
The grounding pipeline adds six stages after ranking: evidence selection, evidence construction with attribution, constrained generation, cross-check and verification (Microsoft Critique), multi-model judging (Microsoft Model Council), and answer generation with citation. Retrieval and ranking produce candidates. They do not produce answers. Both Critique and Model Council are real adversarial review layers your content must survive before being cited.
One measurable consequence: 7% of AI citations go to pages that never appeared in the Google top 10 for the same query. The grounding pipeline follows entity edges, not ranking order. (Source: Volpini synthesis, SEO Week)
Common Crawl releases monthly at large scale and is foundational training data for GPT, LLaMA, and others. It strips schema markup during ingestion. Structured data does not automatically enter training data (Web Commons separately maintains a schema database across the internet). If your content is not inside Common Crawl, you are not in the model’s memory. This is not invisibility. It is absence. Gap detection requires probing the model directly against live API access, not checking a crawl dashboard.
GPTBot (OpenAI), ClaudeBot (Anthropic), Google-Extended, PerplexityBot, and Bingbot all honour robots.txt directives and meta tags. A block on any one removes your site from that system’s grounding pool entirely. Many blocks were added during 2023 as a defensive response and were never revisited. No access equals no visibility.
Freshness is no longer a ranking boost. It is an eligibility gate. Stale pages are filtered out before ranking, at the eligibility stage upstream of the grounding pipeline. A meaningful refresh cadence plus IndexNow keeps evergreen pages eligible. A copyright-year bump does not count as a meaningful update. (Source: Madhavan synthesis, Microsoft SEO Week)
Prose with no entity structure forces the AI to fall back to brute-force vector similarity. Schema markup alone barely moves the needle. WordLift’s benchmark on Vertex AI Vector Search 2.0 found a 29.6% retrieval accuracy lift from navigable entity pages over schema markup alone. Structure is what moves the number, not markup validation. (Source: arXiv:2603.10700, Volpini synthesis) The agent needs traversable entity edges, not just labels.
LLMs do not reliably recall page-level facts from training data alone. A live client audit illustrates this directly: the live house edge on a client’s blackjack page is 0.57%. When probed without live page access, ChatGPT returned 0.50%. Fresh, structured, on-page facts plus IndexNow recrawl are what close that gap and keep the model’s answer correct during buyer evaluation.
Our authority is built on data, not LinkedIn.
AI is a data science problem. Most agencies are not data science companies. We are.
99% of people posting GEO tips on LinkedIn have not run a data study. We have run hundreds.
Here is a relevant one.
11M+ AI citations analysed
The deepest study of its kind
DataForSEO, SurferSEO, SearchAtlas
Proprietary tool kit: Consumer AI visibility tools give you a score. Ours show you what is broken, why it is broken, and what to fix first. Built on your data. An order of magnitude more powerful than anything off the shelf.
"AI does not cite the loudest brand. It cites the most connected one."
The research behind the methodology
11M+ AI citations analysed via the DataForSEO API partnership. Approximately 2,000 hours of research. A follow-on study is running with SearchAtlas and SurferSEO.
Three cosine similarity signals measured per citation: siteFocusScore (site-level topical alignment), siteEmbedding (semantic similarity of domain to query), and pageEmbedding (page-level semantic match). Spearman correlation run against Harmonic Centrality scores to map citation frequency against network position. Ordinal regression on citation rank and cited-vs-not classification against SERP-controlled negatives.
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 programs try to appear in every answer. The correct goal is to become one of the recurring set.
WLDM’s entity audit pipeline is Cloudflare-safe and tiered by model cost. Stage 1: select the 200 to 500 highest-authority entity-bearing URLs via Ahrefs top-pages, avoiding direct origin hits behind bot protection. Stage 2: extract per-URL structured data using JSON-LD-capable methods only. DataForSEO, Ahrefs, and instant_pages return rendered text and are blind to raw <script type=”application/ld+json”> tags. “No schema found” from these tools is a systematic false negative. Always verify with DevTools view-source, Screaming Frog, Firecrawl, or extruct. Stage 3: score and categorise entities by entity_score (page authority × mention count × type bonus). Stage 4: gap analysis from three parallel sources: competitor entity set-diff, Wikidata graph expansion, and ghost citations (entities the model already co-mentions with the brand but the brand does not own as pages). Stage 5: deliver graph.jsonld, entity-page stubs, llms.txt index, edge-list CSV, and a prioritised gap roadmap.
Flagship client validation: 22 agents, 74 tool calls, approximately 7.6 minutes wall-clock time, 3.17M sub-agent tokens. Ahrefs quota consumed: 4,300 of 8,000,000 available units. Scaling to approximately 700 pages requires a count increase, not a redesign.
Implementation, not theory.
Most AI visibility tools track citations. We build them. We show you what is broken, fix it, and build the on-page architecture that moves your citation rate.
We audit whether AI can reach, read, and trust your site against four gates. Every gap gets a priority and a fix in the order that moves the needle first.
Discoverable
Parsable
Groundable
Groundable
The 4 Gates of AI Visibility
The AI crawler has to be able to reach your site. Clean URLs, valid sitemap, IndexNow integration, no inadvertent robots.txt blocks for GPTBot, ClaudeBot, Google-Extended, PerplexityBot, or Bingbot, clean canonical chain. Most enterprise sites have at least one block they do not know about.
An optional llms.txt file (or a dedicated AI info page) can declare canonical entities and point crawlers at the right source of truth. IndexNow fires on publish, meaningful update, and deletion. 50% of clicked newly-indexed Bing URLs originate from IndexNow. (Source: Madhavan synthesis, Microsoft SEO Week)
The AI has to be able to read what it finds. Semantic HTML, valid JSON-LD on every meaningful page present in raw server-rendered HTML (SSR-confirmed, visible to crawlers that do not execute JavaScript), alt text, video transcripts, clean heading hierarchy, content that loads without JavaScript execution.
Critical audit note: DataForSEO, Ahrefs, and instant_pages return rendered text and cannot see raw <script type=”application/ld+json”> tags. “No schema found” from these tools is a systematic false negative, not a measurement. WLDM always verifies structured data with a JSON-LD-capable method: DevTools view-source, Screaming Frog, Firecrawl, or extruct. Fail Gate 02 and every structured data investment becomes invisible.
The AI has to be able to use your content as a citable source. Self-contained passages that answer one question without requiring context from another page, clear provenance on every claim, no contradicting statements on the same page.
This is the gate most well-ranked sites fail. Good content for Google and citable content for AI are not the same document architecture. After retrieval and ranking, the grounding branch adds evidence construction, attribution, and constrained generation, all of which require self-contained, provenance-clear passages.
Flagship client validation: 22 agents, 74 tool calls, approximately 7.6 minutes wall-clock time, 3.17M sub-agent tokens. Ahrefs quota consumed: 4,300 of 8,000,000 available units. Scaling to approximately 700 pages requires a count increase, not a redesign.
The AI has to be able to verify your authority. Consistent @id for each entity on every page (not /author/brie on some pages and /about#brie on others; the same exact URI everywhere), named authors with external verification, multi-surface presence, fact consistency across time.
Gates 3 and 4 are where Microsoft’s grounding pipeline runs Microsoft Critique and Microsoft Model Council, real adversarial review layers. Pass both and you survive the stages most brands never reach. Madhavan’s closing thesis: “In the AI web, visibility is earned by how well machines can trust, understand, and reuse your content — not just rank it.” Trust maps to Gate 04, understand to Gate 02, reuse to Gate 03.
The AI grounding pipeline runs sequentially. Fail Gate 01 and you are disqualified before the AI processes a single word. Fail Gate 02 and structured data you built is invisible to every parser. Fail Gate 04 and you survive retrieval but fail the trust check at the final stage. Most audits find failures at multiple gates. The fix order is not arbitrary. It is sequenced so the highest-leverage gate closes first.
Every entity on your site needs a permanent URI (@id) and you must use that exact URI every time you mention the entity, on every page, character for character. Volpini: “If your entity is named differently in your blog, product pages and Schema.org, the agent can’t resolve them. Three weak signals instead of one strong one.”
WordLift’s benchmark tested three tiers: plain HTML (baseline), HTML + JSON-LD schema markup (modest improvement), HTML + JSON-LD + dedicated entity pages with RDF links (+29.6% retrieval accuracy). Schema markup alone barely moved the needle. Navigable entity pages did. (Source: arXiv:2603.10700)
The highest-leverage single on-page change is @id consistency. The exact same URI string for the same entity on every page, site-wide. WLDM recommends a CI test that flags inconsistent @ids on every PR. It is nearly free and immediately measurable.
Five steps. (1) Identify the 50 to 500 things the site is about across products, people, places, and concepts. Deliverable: a flat list grouped by type. (2) Assign a permanent URI pattern (/entity/[slug]) and hold to it. An @id is a primary key; changing it is a migration. (3) Create one entity page per core entity with @type, stable @id, name, alternateName, description, sameAs links to external authorities, and subjectOf back-edges to every content page about it. (4) Find Wikidata Q-numbers for every globally-notable entity and add them to sameAs — proprietary entities that do not appear on Wikidata stay as local URIs. (5) Make @id consistent site-wide.
Entity pages and content pages are two different things that target different intents. The entity page tells anyone (human or AI) what a thing is. The content page tells a user how to use it. Hub-and-spoke. They do not cannibalise each other. It is the pattern Wikipedia uses.
WLDM runs 45 live prompts about the client’s brand using the ChatGPT API with force_web_search=true and Confident 3x mode. Prompts cover category evaluation, brand comparison, feature-specific, use case, and objection-handling queries. Every model response is scored against ground truth from the client’s own pages. Output: Hallucination Risk Score per entity: which facts the model approximates, which it gets wrong, and the specific on-page gap causing each inaccuracy. Fix-priority order is set by commercial query volume and gap severity. Fragment-level data-nosnippet directives give precision control: pages stay eligible for grounding while specific incorrect fragments are excluded from extraction.
GEO owns the on-page layer: crawler compliance, Common Crawl indexation eligibility, entity schema construction, @id consistency, and hallucination correction. Off-page citation placement (co-citation cluster engineering, Harmonic Centrality positioning, and the authority network programme) belongs to the AI Citations service.
Phase 4 of the engagement identifies the gap and hands the outreach target list to AI Citations. The on-page work creates the eligibility. The off-page work does the placement.
Named everywhere. Cited nowhere.
On the world’s largest crypto sportsbook, AI assistants named the brand in 89% of responses. The brand’s own site was cited in 11% of them. Affiliates and review sites controlled the rest.
That is not a visibility problem. That is a source problem. The fix is not more content. It is making your own site the source AI trusts.
Brand mention rate in AI search (Stake)
How often AI cited the brand’s own website
Own-domain citations on category queries (“best crypto sportsbook”) at baseline
How we measured it
WLDM ran 45 live prompts about the client across query intent types: category evaluation, brand comparison, feature-specific, use-case, and objection-handling. Every prompt executed via the ChatGPT API with force_web_search=true (live web access, not training-data recall) and Confident 3x mode (a citation must appear in three consecutive independent API calls to be counted as a stable citation).
Canada proxy removes localisation variance. This is direct API access, not scraping. Results are consistent with what buyers actually see.
Brand name queries: 89% mention rate. Own-domain citation rate: 11%. Category queries: 0% own-domain citations at baseline. Affiliates and review sites controlled the answer entirely.
The one prompt where the client’s own domain earned a stable citation was on a page with a dedicated factual entity page behind it. The page earned the citation on visible HTML and og: tags, despite broken page-level schema on the same template — which only sharpens the upside of fixing the full entity layer across the site.
The original audit reported zero structured data. That finding was retracted. It was a tool-blindness trap. DataForSEO, Ahrefs, and instant_pages return rendered text and cannot see raw script tags. The real finding: the client ships a sophisticated entity graph undermined by two critical bugs.
Bug 1 — core nodes emitted as escaped JSON strings: On the majority of templates the WebSite node is double-encoded as a string literal inside a JSON-LD array, making the entire node invisible to every parser. That dead node holds the entity sponsor roster, the site-level about link to Wikipedia, and author references. The correct serialization already exists on some templates. The fix: stop double-encoding. One change resurrects the whole graph.
Bug 2 — fragment @ids never consolidate: Every entity uses a bare fragment, which resolves against the document URL, so the same entity gets a different absolute @id on every page. Volpini’s first rule, broken structurally. Fix: mint absolute reused @ids.
217 entities reconciled by declaration status: 124 declared-valid, 22 declared-bare, 21 declared-broken, 31 mentioned-only, 19 undeclared. Wikidata declared on 0 today. 89 of 217 QIDs resolved and ready to drop into sameAs.
Discoverable: amber. Parsable: red (schema present but broken, visible only to view-source, invisible to standard audit tools). Groundable: red (hub pages near-empty of extractable text). Trustworthy: red (no Wikidata grounding; on brand legitimacy queries the AI volunteers regulatory caveats).
The GEO entity work is what moves the 11% own-domain citation rate. Re-baseline is scheduled at Week 13 of the engagement.
We will show you exactly how AI sees your brand, and where it is getting it wrong.
Every brand we audit finds a gap. The only question is whether you find it before your buyers do.
We audit your site against the 4 gates every AI grounding pipeline runs through: Discoverable, Parsable, Groundable, Trustworthy. Fail any gate and you are disqualified from later stages, regardless of how well your content is written. We map exactly where you are losing citations and give you a priority-ordered implementation roadmap that moves your citation rate first.
What we are delivering: AI citing your brand accurately, from your own site, on the queries that matter to your buyers.
What the audit covers
1. Target Prompt Database
100 prompts mapped across intent types: category evaluation, brand comparison, feature-specific, use-case, objection-handling, and competitive displacement. The database drives the baseline measurement and every subsequent re-baseline at quarterly intervals.
2. AI Crawler Compliance Audit
robots.txt and meta directive check for GPTBot, ClaudeBot, Google-Extended, PerplexityBot, and Bingbot. Every block removing you from a grounding pool is identified and prioritised for removal. Includes llms.txt file assessment and AI info page recommendation where appropriate.
3. Industry Corpus Inclusion Scorecard
Common Crawl presence check and entity recognition across foundation model training data. Scored by entity type: brand, product, people, concepts, categories. Identifies which entities the models have baseline training recall on and which are absent.
4. Training Data Coverage Report
Per-entity model recall scoring across the 100-prompt database. Documents which entities the model knows, which it approximates with incomplete information, and which it has no baseline on at all.
5. Training Data Gap Map
Entities where model recall is absent or materially inaccurate versus ground truth from your own pages. Priority-ordered by commercial query volume and buyer-evaluation impact.
6. AI Visibility Monitoring Dashboard
Ongoing own-domain citation rate tracking against the Target Prompt Database. Quarterly trend reported against competitor citation rates. Built on direct ChatGPT API access, not scraping. Results are consistent with what buyers see.
7. Competitive Knowledge Graph and Topic Gap Analysis
Competitor entity inventory compared against yours. Surfaces the concepts, products, and people your competitors have entity pages for that you do not, and which of those gaps are costing you citation share on commercial queries.
8. Query Fan-Out Prompt Database
Your top commercial queries expanded across five grounding lanes: aesthetic, popularity, engagement, reviews, and freshness. One query becomes five rewritten forms across these lanes. Citation coverage requires presence in every lane — coverage on one means four paths missed. (Source: Madhavan synthesis, Microsoft SEO Week)
9. Hallucination Risk Assessment
45-prompt live API benchmark per brand. Every model response scored against on-page ground truth from your own pages. Hallucination Risk Score per entity. Fix-priority list mapped to remediation steps.
10. Citation Cluster and Centrality Map
How your brand connects (or does not) to the network of sources AI treats as authoritative in your category. This is the on-page component of what moves the own-domain citation rate.
11. Co-Citation Gap Report and Outreach Target List
Which high-centrality sources co-cite your competitors but not you. This is the handoff list for the AI Citations off-page programme. The on-page work creates the eligibility. The off-page work does the placement.
