AI SearchSeptember 9, 202612 min readBy the Kiwi senior team

73% of enterprises rank on Google page one and get zero AI citations — the architecture difference

Ranking and being quoted are different achievements. The checked 2026 evidence on what makes B2B content citable by AI answers, which page types earn citations and which never will, the honest schema question, and how to measure presence instead of sessions.

In this guide

Ranking first and being quoted are different achievements
What citable actually means — three properties, not one tactic
Why B2B is structurally harder than B2C
The formatting rules that cost nothing
Which page types produce citations — and which never will
The schema question, answered honestly
Entity clarity — the off-site half nobody fixes
Measure presence, not sessions — and the 90-day plan
Questions
Sources

Ranking first and being quoted are different achievements

The most uncomfortable number in 2026 B2B search is this: Onely's analysis found that over 73% of enterprises have zero mentions in AI-generated responses despite ranking on Google page one. The two outcomes are produced by different mechanisms. Ranking rewards topical coverage across a whole document. Citation rewards a specific, extractable assertion inside it — and the retrieval layer does not read your page the way a person does. It splits the page into chunks and scores each chunk independently.

That is why a page can hold position one for a query and still be useless to a generative system. There is no single safe sentence for the model to lift. The gap is not effort or budget; it is architecture. Most B2B content is written to reassure a human who has already arrived, not to be extracted by a system deciding what to quote. Fixing that is a writing and structure problem, and it is far more tractable than the AI-visibility tooling market would have you believe.

73% of enterprises have zero AI mentions despite Google page-one rankings (Onely, 2026).
Ranking scores the whole document; citation scores one extractable passage inside it.
Retrieval chunks your page and scores chunks independently — there is no 'page rank' inside an AI answer.

What citable actually means — three properties, not one tactic

Citable content has three properties, and a passage needs all three. It is self-contained: paste any paragraph into a blank document with no heading and no context, and a stranger can still tell what it is about, who claims it, and what it claims. A paragraph that opens with 'This means that' or 'As a result' fails immediately. It is attributable: your name is attached to the claim. 'Industry research suggests' hands the citation to the research body; 'In our 2026 review of 180 UK migrations, we found' makes your organization the origin of the fact.

It is verifiable: the method is visible. You do not need a peer-reviewed study, you need to say where a number came from — the sample, the period, the definition. 'Across 180 projects delivered between January 2024 and December 2025' is a method statement, and it is what separates citable content from a figure a system will quietly drop. The reliable block shape follows from these three properties: a question-shaped heading, a direct answer in the first sentence, the specific evidence, the conditions under which it holds, and the attribution. Written well, that block runs 60 to 100 words and looks unremarkable to a human reader — which is the point.

Self-contained: survives being cut out with no heading or context.
Attributable: your name attached to the claim, not 'industry research suggests'.
Verifiable: the sample, period, and definition are visible in the sentence.

Why B2B is structurally harder than B2C

Consumer content answers questions with public, checkable facts — opening hours, dimensions, prices. B2B content answers questions whose real answer is 'it depends', and the sales instinct is to protect that answer until a call is booked. That instinct is exactly what strips a page of citable content. As the practitioners who study this put it bluntly: machines cannot cite a discovery call. Every 'contact us for pricing' and every hedged 'results vary' is a passage a synthesis engine has nothing safe to extract from.

The fix is not to give away the whole consultation. It is to state the dependency as a boundary rather than a hedge. Not 'results vary considerably', but 'this holds for estates under 250 seats; above that, licensing changes the arithmetic'. A boundary is a fact, and facts are what citable content is made of. A hedge is an absence. B2B teams that understand this distinction can keep the commercial gating they need while still publishing the specific, bounded claims that earn a citation.

B2B answers are 'it depends'; the sales instinct protects the answer until a call is booked.
Machines cannot cite a discovery call — 'price on request' yields nothing extractable.
State dependencies as boundaries ('holds under 250 seats'), not hedges ('results vary').

The formatting rules that cost nothing

Structure does not create substance, but bad structure hides good substance and stops it becoming citable. The highest-yield edit available on almost any B2B page is the heading: rewrite 'Our Approach' as the question the buyer actually typed. A question-shaped H2 is the strongest chunk-boundary signal you control, and it keeps the answer beneath it inside one coherent, retrievable block. Front-load lists with the noun or the number — 'Six to nine weeks for a 40-user rollout' gives the model a citable line per item, where a verb-first bullet that runs three lines does not.

The rest is mechanical and mostly free. Do not bury the answer in paragraph four; move it to sentence one. Publish a price band and its drivers instead of 'price on request'. Repeat any figure that lives inside an image as HTML text, because text retrieval cannot read it otherwise. Server-render the main content, since a client-side-rendered body may fetch as an empty page. Publish findings as an indexable page rather than a gated PDF that is never crawled. Show published and reviewed dates. Convert prose comparisons into tables, which extract row-wise. None of this is exotic; it is the difference between a page that is retrievable and a page that is quotable.

Question-shaped H2 is the highest-yield citability edit on almost any B2B page.
Front-load lists with the noun or number; move the answer to sentence one.
Server-render content, publish price bands, repeat image figures as text, convert comparisons to tables.

Which page types produce citations — and which never will

Not every page is worth optimizing, and the citation-potential ranking is consistent across the 2026 guides. The reliable producers are pricing and cost explainers, comparison and 'X versus Y' pages, definitional guides, standards and compliance explainers, benchmark and survey write-ups, technical documentation, and genuinely specific FAQ pages. They earn citations because they are shaped like answers to questions people actually ask a model.

The near-zero end is where most B2B content budget is currently pointed. Generic service pages score low unless you split out a specifics section. Thought-leadership essays score very low unless you add original data. Homepages, careers pages, and awards pages are near-zero and should be left out of the program entirely. A case study is salvageable — medium potential once rewritten with named scope, duration, and metrics — but in its default 'we partnered and achieved great results' form it is not citable. The audit question is not 'is this page good'; it is 'does this page contain a sentence a model can safely quote'.

Very high citation potential: pricing/cost explainers, comparison and 'vs' pages.
High: definitional guides, standards/compliance, benchmark and survey write-ups, technical docs.
Near zero: homepage, careers, awards, and un-rewritten thought-leadership essays.

The schema question, answered honestly

Every AI-visibility pitch eventually lands on schema, and the honest answer is more muted than the marketing. One vendor, Amicited, reports that FAQ schema improves citation probability by 28 to 40 percent — but that figure comes without a disclosed sample or methodology, so treat it as a vendor claim. The stronger independent evidence points the other way: a controlled Ahrefs test tracking 1,885 pages that added JSON-LD against matched controls found no meaningful citation lift on ChatGPT, Google AI Mode, or AI Overviews.

The practitioners who write the best citable-content guides agree on the correct framing: schema does not make content citable on its own. It removes ambiguity about what a page is and who published it, which helps a parsing layer, but no amount of markup makes a vague page quotable. So do the work that matters — Organization markup with sameAs links to verified profiles, Article or FAQPage markup matched to the visible text, Product or Service markup where it applies — but do it after the writing, not instead of it. Structure and evidence convert a retrieved chunk into a named citation; schema only makes the chunk easier to interpret.

Amicited claims FAQ schema lifts citation 28-40% — a vendor figure with no disclosed method.
Ahrefs controlled test (1,885 pages): no meaningful citation lift from adding JSON-LD.
Do schema after the writing: it removes ambiguity about who is speaking, it does not create a quotable claim.

Entity clarity — the off-site half nobody fixes

Citation is only half on-site. The other half is whether a model can resolve you to one coherent entity. Your legal name, trading name, address, sector descriptors, and service names should be identical across your website, Companies House, LinkedIn, review platforms, directories, and every profile you have ever created. Inconsistent descriptions split your entity into several weak ones; consistent descriptions merge into a single resolvable organization a model can reason about. That resolution is the precondition for treating anything you publish as citable.

The agentic-buying research makes the stakes concrete. MarTech's analysis of AI agents handling B2B procurement concludes that market share in an agent-mediated environment is determined by technical data architecture, and that the three signals that function as the agent-era equivalent of domain authority are structured data markup, semantic schemas that define what a product does and does not do, and citations from high-authority independent publishers. When an agent is doing the shortlisting, your brand's first impression is a schema tag and a consistent entity, not a homepage hero image. Off-site consistency is not branding housekeeping; it is discoverability infrastructure.

Name, address, sector, and service descriptors must be identical across site, registry, LinkedIn, reviews, and directories.
Inconsistent descriptions split your entity into several weak ones; consistent ones merge into a resolvable organization.
Agent-era authority = structured markup + semantic schema + citations from high-authority independent publishers (MarTech).

Measure presence, not sessions — and the 90-day plan

You cannot manage what you measure with the wrong instrument, and session analytics is the wrong instrument for citation. The measurement that works is presence: build a fixed set of 40 to 60 buyer questions, run them against each assistant on a schedule, and record whether you appeared, whether you were cited with a link, and which page was cited. The appearance rate is the headline number. MarketScale's field data gives a sense of the baseline — an overall citation rate around 25%, a 7.4 percentage-point spread across platforms, and a citation-to-mention ratio of 0.75 on Perplexity — and 89% of B2B buyers now use AI tools during research, so the surface worth measuring is already where your buyers are.

Sequence the work over 90 days rather than buying a dashboard. Days 1 to 15: confirm crawler access for every relevant agent, verify key pages render server-side, fix anything returning a 403 to a legitimate crawler, add Organization schema with sameAs links, and take a baseline reading of your 40 to 60 questions. Then rewrite the very-high-yield page types first — pricing explainers and comparison pages — scoring each against a simple 100-point rubric: answer structure 30, evidence and specificity 25, entity clarity 20, technical access 15, freshness 10. Expect six to twelve weeks for the first movement on rewritten pages. And keep the SEO foundation intact: 44.6% of B2B revenue is still attributed to organic search, so this work is additive to ranking, not a replacement for it.

Measure presence: a fixed set of 40-60 buyer questions, run on a schedule, recording appearance and cited page.
Baseline: ~25% overall citation rate, 7.4pp platform spread (MarketScale); 89% of buyers use AI in research.
90 days: crawler access + baseline + Organization schema first, then rewrite pricing and comparison pages.
SEO is still the foundation — 44.6% of B2B revenue is attributed to organic search (SeoProfy).

Frequently asked questions

Why do we rank on Google page one but never get cited by AI?

Because ranking and citation are produced by different mechanisms. Ranking rewards topical coverage across a whole document; citation rewards one specific, extractable assertion inside it. The retrieval layer chunks your page and scores each chunk independently, so a page can hold position one and still contain no sentence a model can safely quote. Onely's 2026 analysis found more than 73% of enterprises have zero AI mentions despite page-one rankings — the gap is content architecture, not search performance.

What actually makes content citable by an AI answer engine?

Three properties, all required. Self-contained: the passage survives being cut out with no heading or context. Attributable: your name is attached to the claim rather than 'industry research suggests'. Verifiable: the method is visible — the sample, period, and definition. The reliable block shape is a question-shaped heading, the direct answer in the first sentence, specific evidence, a conditions clause, and attribution, in roughly 60 to 100 words.

Does adding schema markup get my pages cited?

Not on its own. A vendor (Amicited) claims FAQ schema improves citation probability by 28-40%, but that figure has no disclosed methodology, and a controlled Ahrefs test of 1,885 pages that added JSON-LD found no meaningful citation lift on ChatGPT, Google AI Mode, or AI Overviews. Schema removes ambiguity about what a page is and who published it, so it is worth doing — but after the writing. Structure and evidence convert a retrieved chunk into a citation; markup only makes it easier to interpret.

Which pages should I rewrite first for AI citation?

The ones shaped like answers to questions people ask a model: pricing and cost explainers and comparison or 'versus' pages are the highest-yield, then definitional guides, standards and compliance explainers, benchmark and survey write-ups, and technical documentation. Generic service pages score low unless you split out a specifics section, thought-leadership essays score very low without original data, and homepages, careers, and awards pages are near-zero — leave them out of the program.

How do I measure AI citation visibility?

Measure presence, not sessions. Build a fixed set of 40 to 60 buyer questions, run them against each assistant on a schedule, and record whether you appeared, whether you were cited with a link, and which page was cited — the appearance rate is the headline number. Expect six to twelve weeks for first movement on rewritten pages. Keep SEO running alongside it: 44.6% of B2B revenue is still attributed to organic search, so citability is additive to ranking, not a replacement.

Sources

Figures cited above are drawn from the linked publications and are the responsibility of their sources; we date and scope them rather than presenting them as universal guarantees.