By Mark Buraga, Independent SEO Consultant, Growth Engine PH. Last updated 7 July 2026.

An Ahrefs study tracked 1,885 pages that newly added schema markup and matched them against 4,000 control pages. The result: essentially no lift in AI citations across ChatGPT, AI Mode, or AI Overviews. Meanwhile, firms with mediocre markup get cited in AI answers every day. The difference between those two groups was never the code. It was entity coherence, and it is the audit almost nobody runs.

The industry read that study as a verdict on schema. It is not. It is evidence that everyone has been auditing the label instead of the box. What AI engines need before they can cite you, mention you, or even describe you accurately is one coherent, corroborated identity: the same name, the same claims, the same facts about who you are, everywhere you exist. Entity coherence fails silently, which is why it goes unaudited. Every page loads. Every rich result validates. And the machine still cannot assemble who you are.

This is the analysis of that missing audit: what entity coherence actually is, why it moved from nice-to-have to load-bearing, what incoherence looks like in the wild, and the exact checks to run. It ends with the audit our own site failed last week, because first-hand embarrassment is more useful than theory.

What Is Entity Coherence?

section-1-assembly

Entity coherence is the degree to which every machine-readable trace of your business tells the same story: one canonical name, one set of facts (what you do, where you are, who your people are), and independent corroboration of those facts on surfaces you do not control. When coherence is high, a search engine or language model can assemble your identity into a single confident entity. When it is low, the machine holds fragments that may or may not be the same business, and it hedges, blends, or guesses.

The distinction that matters: entity coherence is not schema markup. Schema is one way to state your identity. Coherence is whether the identity itself holds together across your site, your directories, your professional profiles, your coverage, and everyone else’s description of you. Markup describes; coherence corroborates.

For boutique professional services firms, the entity is rarely just the firm. It is the firm plus its named partners plus its practice areas, and the connections between them. A specialist firm whose expertise lives in its people has an entity graph, whether or not anyone maintains it.

Why Was Entity Coherence Skippable Before AI Search?

section-2-name-drift

Classic ranking was forgiving of identity mess. Google matched queries to pages, and a page could rank on its own merits even if your firm’s name appeared three different ways across the web. The entity layer existed, Google has run its Knowledge Graph since 2012, but for most small firms it was decorative. You could skip the audit and lose almost nothing.

AI search removed that forgiveness, for one structural reason: answer engines do not retrieve pages, they synthesize descriptions. When ChatGPT, Gemini, Perplexity, or Google’s AI Mode answers “best boutique audit firms for cross-border work,” it is not ranking your page. It is assembling what it believes about entities and deciding which ones it trusts enough to name. Assembly requires coherence. A model that holds three conflicting fragments of your firm does not average them into a flattering summary. It either drops you from the answer or fills the gaps with statistically plausible fiction.

The fiction part is measured. A 2023 Cureus study found 47 percent of references produced by ChatGPT-3.5 were entirely fabricated and another 46 percent carried wrong details, leaving 7 percent fully correct. Models fill identity gaps the same way they fill citation gaps: confidently. If engines cannot verify who you are, they guess, and their guess becomes what your next referral reads. With AI Overviews now appearing on roughly 25 percent of 21.9 million studied queries and climbing on informational searches, the surface area for that guess grows monthly.

Does Schema Markup Earn AI Citations? (The Debate Is a Category Error)

section-3-audit-circuit

The 2026 argument splits practitioners into two camps. Camp one sells JSON-LD as an AI-citation tactic: mark everything up, win the answers. Camp two waves the Ahrefs null result and calls schema dead weight. Both camps are looking at syntax.

Take the strongest version of the skeptic case seriously, because it is good evidence. The Ahrefs design was clean: pages that newly added JSON-LD, matched controls, three AI surfaces, no measurable citation lift. Google’s own Gary Illyes said the quiet part at an SEOFOMO meetup: it “would be lovely” if Google did not have to rely on schema at all, because ideally systems “would simply understand the page.” Third-party LLMs fetching your pages at query time do not even parse JSON-LD as structured data; they tokenize it as plain text alongside everything else.

All of that is true, and none of it says entities stopped mattering. It says markup describing an incoherent identity is a label on an empty box. The sites earning citations in that same study period were the ones whose identity was already assembled: consistent naming, corroborated facts, real footprints on independent surfaces. Coherence works without markup. Markup without coherence does nothing. That asymmetry is the whole finding, and it points the audit one layer down.

Schema still belongs in the stack. Google’s first-party surfaces consume it as runtime context, and it remains the cheapest way to state your facts unambiguously. The correction is sequencing: schema encodes an identity. It cannot create one.

What Does Entity Incoherence Look Like? (Six Failure Modes)

These are the patterns we find in audits of specialist firms, ordered by how often they appear:

  1. Name-variant drift. The legal name, the brand name, and the abbreviation all circulate, and no single form dominates. The website says one thing, the bar or professional registry says another, LinkedIn says a third. Each variant splits your corroboration three ways.
  2. The about-page gap. No page states, in plain prose, what the firm is, who runs it, where it operates, and what it specializes in. Identity facts are scattered across a footer, a PDF brochure, and a partner’s bio. Humans infer; machines do not.
  3. Orphan experts. Partners with real reputations, credentials, and publications, and no page that ties them to the firm entity. The expertise the firm sells is invisible to the graph that decides whether the firm is an authority.
  4. Contact and location drift. Three phone numbers across the site, a fourth on the business profile, an old address in a directory. Classic NAP inconsistency, except the consequence is no longer a weaker map ranking. It is a synthesis engine treating your locations as evidence of two different businesses.
  5. Self-referential identity. Everything the web knows about you is published by you. No directories, no professional listings, no coverage, no third-party description. Coherent, but uncorroborated, and models weigh corroboration heavily.
  6. Description conflict. Your site says “boutique advisory consultancy,” a directory says “accounting firm,” an old press mention says “tax preparer.” None are false. Together they blur the one thing a model must resolve: what are you, exactly?

Notice that no SEO tool flags any of these. Crawlers report on pages. Coherence lives between pages, and between your site and the rest of the web.

How Do You Audit Entity Coherence?

The audit that catches the six failure modes. It takes a focused afternoon for a typical specialist firm, and none of it requires a developer:

#CheckPass looks like
1Canonical name censusOne primary name form chosen; every owned surface (site, profiles, directories) uses it exactly; variants documented and mapped, not scattered
2Identity statementOne about page that states what you are, who leads it, where you operate, and your specialisms in plain prose a machine can quote
3Person-entity pagesEach named expert has a page: credentials, role, practice areas, linked both ways with the firm
4Contact and location consistencyOne primary phone, address, and email set, identical across site, business profiles, and directories
5Corroboration censusList every independent surface that describes you (registries, directories, associations, coverage). Fewer than five is a red flag for an established firm
6Description alignmentThird-party descriptions match your identity statement in substance; conflicting ones get correction requests
7Cross-linkingYour profiles point to your site; your site points back to your profiles; the loop is closed in both directions
8The assembly testAsk three AI engines who you are, unbranded and branded. Compare their answer to your identity statement. Every error is a coherence gap with a to-do attached

Then, and only then, encode the result: Organization and Person schema restating what the audit made true, with corroborating profiles referenced from the markup. The markup now describes a box with something in it.

Run check 8 quarterly. It is the only check that measures the outcome rather than the inputs, and it doubles as your before-and-after evidence. We run a version of it weekly for clients, prompt by prompt, engine by engine, because testing AI citation is the feedback loop the whole discipline runs on.

The Audit Our Own Site Failed

Last week we ran our own practice through the same process we sell. The technical layer came back almost embarrassing in how clean it was: crawler access 100 out of 100, every schema type server-rendered, zero critical issues. A site any auditor would wave through.

Then we ran the AI-visibility layer: nine unbranded questions a real buyer would ask. Zero citations. Not one engine named us for anything a stranger would search.

Our first hypothesis was a crawl problem. Guide the bots better, check the server logs, maybe the blog is not being discovered. So we probed live search results instead of guessing. The blog was crawled. It was indexed. It even ranked, for brand terms where no one competes. The hypothesis died in an afternoon.

What was actually missing was everything off the page: two referring domains, name signals scattered, an identity story that lived only on surfaces we control. The machine had crawled everything and assembled nothing, because nothing independent corroborated who we were. Every tool we own reported green while the audit that mattered failed silently. That finding rewrote our own quarter: less content, more corroboration. The same medicine we prescribe.

The honest contrast: a client of ours, a boutique legal and accountancy firm, spent six months on the unglamorous half of this work, one name form, consistent contact facts, per-page identity, corroborated claims. This week our tracker measured it cited or mentioned on 62 percent of 21 watched prompts. Same discipline, opposite result, and the variable was never markup.

What Should You Ask an SEO Vendor About Entity Work?

If you are a managing partner reading an SEO proposal: look for the entity work. Crawl fixes, content calendars, and schema line-items are all legitimate, and none of them touch the layer that decides whether an answer engine can say your name with confidence. Ask the bidder what they would do about name variants, about partner entities, about corroboration. Blank looks are diagnostic.

If you run marketing for a specialist firm: the audit above is deliberately runnable without a vendor. Checks 1, 2, and 4 are an afternoon with a spreadsheet. Check 8 costs nothing but honesty.

And if you are a practitioner: stop letting the schema debate set the agenda. Both camps are right about the markup and silent about the entity. The gap between them is where the results are.

Schema tells engines what you claim. Coherence lets them believe it. Audit the second one first.

FAQ

What is entity coherence in SEO? Entity coherence is how consistently your business’s identity (name, facts, people, locations, specialisms) is stated and corroborated across your website and independent surfaces. High coherence lets search and AI engines assemble you into one confident entity they can cite; low coherence makes them hedge, blend you with others, or guess.

Is entity coherence the same as schema markup? No. Schema markup is one way to state your identity in machine-readable form. Entity coherence is whether the identity itself is consistent and corroborated. An Ahrefs study of 1,885 pages found markup alone produced no lift in AI citations; coherence is the layer that does the work markup gets credit for.

Is this just NAP consistency renamed? NAP (name, address, phone) consistency is one row of the audit. Entity coherence also covers name variants, the about-page identity statement, person entities for your named experts, third-party description alignment, and corroboration by independent sources. Local citations never had to survive a language model’s synthesis step.

How do I test my firm’s entity coherence? Run the assembly test: ask ChatGPT, Gemini, and Perplexity who your firm is and what it does, both by name and through unbranded questions a buyer would ask. Compare the answers against your own about page. Every error, blend, or blank is a coherence gap you can trace to a specific surface.

Does schema markup still matter for AI search? Yes, with corrected expectations. Google’s own AI surfaces use structured data as context, and markup is the cheapest way to state facts unambiguously. But it encodes an identity rather than creating one. Sequence it after the coherence audit, not instead of it.

Why would an AI engine describe my business incorrectly? Language models fill gaps with statistically plausible text. A 2023 Cureus study measured 47 percent fully fabricated references from ChatGPT-3.5. If your identity signals are thin or conflicting, the model’s description of you is partly its best guess, and it delivers that guess with full confidence.

How long does entity coherence work take to show results? The audit takes an afternoon; the fixes range from same-day (contact consistency, about page) to months (corroboration on independent surfaces). In our client work, measurable AI-visibility movement followed roughly six months of consistent identity plus corroboration work, tracked weekly per prompt and engine.

Sources

Related reading