What drives LLM citation rank — and where the engines diverge

Observational, confound-controlled study: 138,186 source citations, 4 LLMs, 3 company datasets, full page re-audit + query-relevance + external authority. Associational, not causal.

How to read this report

Key terms — in plain English

Odds-ratio (OR)

How much a page feature multiplies the odds of landing in the top 3. 1.0 = no effect; 1.2 = 20% better odds; 0.8 = 20% worse. Nearly everything here is 0.9–1.1 (small).

Lift

The plain version of OR — how many times more likely a page is to support a top-3 company with a feature vs without it.

AUC

How well a model predicts top-3 ranking: 0.5 = a coin flip (useless), 1.0 = perfect. Our page-feature models sit at 0.49–0.61 — features barely predict rank.

Significance (q<.05, ✱)

The result is unlikely to be chance, after correcting for testing many features at once. More stars = more confidence.

Content vs boilerplate schema

Structured-data tags in a page's code. Boilerplate = auto-generated tags every CMS emits (WebSite, Organization). Content = meaningful tags describing the page (FAQ, Article, LocalBusiness, Review).

Answerability

Whether a page directly answers the question — definitions, examples, a clear conclusion — rather than just marketing copy.

Brand / company effect

The part of ranking explained by which company it is, independent of its pages — a company's stable tendency to be cited after page features are removed.

Popularity (proxy)

How many distinct questions a company shows up for at all — a rough, in-dataset stand-in for prominence.

External authority

Classic SEO authority — how many other sites link to a domain (Majestic Million rank). An independent signal, unlike "popularity".

Eligibility vs rank

Eligibility = being the kind of page that can be cited at all; rank = how high you place once eligible. Most signals affect eligibility, not rank.

Cohorts

Cited = pages LLMs cited · sibling = other (uncited) pages on those same sites · Majestic / CC-web = random established / typical web pages, for comparison.

Cross-engine overlap (Jaccard)

How much two engines' top-3 lists share: 0 = nothing in common, 1 = identical. Ours ≈ 0.09.

Consensus core

The handful of brands that all four engines put in the top 3 for the same question — the thin "head" they agree on despite disagreeing on everything below it.

Transfer

If a company is top-3 on one engine, how often it's also top-3 on another. Low transfer ⇒ winning on ChatGPT doesn't get you onto Perplexity.

Localized vs national query

Localized = the prompt names a UK city ("…in Manchester"). National = it names no place ("…in the UK").

National incumbent

For the localization section: a company cited across ≥3 different cities and among the 15 most-cited such firms (national scale + footprint) — excludes the client itself and single-city local brokers.

Order of magnitude

"≈10× or more." When we say the company effect is an order of magnitude larger than page features, one is in the tens-of-percent range and the other in the multiples range.

Cluster-robust CI / Bayesian HDI

Two stricter ways to draw the error bars — one corrects for the same company appearing many times, the other re-estimates from scratch. Both agree with the standard result here.

Correlation (Spearman ρ)

Strength & direction of a relationship, −1 to +1; 0 = no relationship.

Executive summary

What matters most — 10 stats for ranking on LLMs

1. Brand identity is ~everything. A company's stable identity outweighs every page feature by more than an order of magnitude (hierarchical model: company SD ~3 vs feature coefs <0.1) — and that identity is not domain authority, on-page craft, or query-relevance.
2. Backlinks / domain authority do NOT move LLM rank. External authority (Majestic) predicts top-3 at AUC ≈ 0.50 (coin flip). You cannot link-build your way to citations.
3. Engines diverge on the tail, agree on the head. Only ~9% full top-3 overlap, but the same brand is #1 on ≥2 engines 37% of the time — a thin consensus core of big brands, divergent everywhere below.
4. Localization helps locals only in some cities — not as a rule. Naming a city doesn't shrink national-incumbent share on average (17% localized vs 12% national). But it's sharply city-specific: incumbents hold 0% in Glasgow and under 15% in 4 of 6 cities, vs ~27% in Leeds.
5. Eligibility = answerability + content schema, not boilerplate. Cited pages are ~3× the typical web on answerability (74% vs CC-web 18%) and content schema (46% vs 19%) — LocalBusiness/FAQ/Article/Review. But boilerplate schema (WebSite/Org) is everywhere (~40–60%) and doesn't discriminate, and none of it moves rank once cited.
6. Page speed is a near-non-factor (OR ≈ 1.0; small per-engine differences only).
7. Stats beside claims don't help citation here — cited pages had lower stat-density than random web. Contrarian vs Princeton/Citation-Absorption.
8. Q&A/FAQ helps a little, on every engine (+1–4%) — neither the +25% nor the −5.7% the literature claims.
9. No freshness cliff — citation rate is flat from <1mo to >12mo old content.
10. Readability is the one consistent on-page positive — robustly + on OpenAI/Perplexity, − on Anthropic.
11. Don't keyword-stuff — title↔body keyword matching is negative on Gemini (0.73×), echoing Princeton.

The playbook grounded in the findings below

Per-engine fingerprints what matters on each

Tilts are small (±~10%) — brand prominence dominates all four — but these are the levers that move the margins once you're in the consideration set.

1 · Per-LLM driver matrix logistic, top-3, FDR-corrected

Takeaway: Read this as "how much does each page feature change the odds of ranking top-3, per engine." Everything sits near 1.0 (±10%) — page features barely move rank. The few real signals: completeness & readability help (esp. ChatGPT); keyword-stuffing hurts (Gemini).

Odds-ratio per +1 SD / presence, controlling for company effect, site, niche. *q<.05 **q<.01 ***q<.001. Effects are small (±10%) — read directions, not dials. Two cells are artifacts, not signals: stat_density's anthropic 2.09× is a heavy-tail outlier (robust lift ≈ 1.0×), and the presence/completeness split is collinearity (r=0.835) — read their combined effect (schema-rich lift ≈ 1.0 = neutral for rank), not the individual signs. Schema is an eligibility signal (§8), not a rank lever.

Model fit: openai AUC 0.601 · anthropic AUC 0.491 · gemini AUC 0.539 · perplexity AUC 0.534 — all near 0.5–0.6 (features weakly predictive).

2 · Prominence, authority & relevance marginal AUC

Takeaway: Nothing page-level beats simply knowing the company. Adding all page features + query-relevance + backlink authority doesn't predict rank better than the company effect alone — and backlinks alone are a coin flip.

Does anything page-level beat just knowing the company? Cross-validated AUC by signal set:

On every engine, the company effect alone ≥ everything combined; relevance ORs 1.02–1.10 (negligible).

Is it real authority, or circular? — external check

My "popularity" was an in-dataset appearance count (partly circular). The honest test joins cited domains to Majestic Million (backlink/referring-subnet rank) — a true external authority signal. 77% of cited domains are in the top-1M.

3 · Cross-LLM divergence

Takeaway: The four engines mostly disagree on the full top-3 (~9% overlap) — but they do share a thin head of recognized brands (consensus core below). Optimise per engine; expect agreement only on the biggest names.

Do drivers differ by engine? (interaction tests)

Source ecology — what each engine cites (role share)

Self = own page. Perplexity/OpenAI lean on own pages (~half); Anthropic on third-party review/comparison. Only ~9% top-3 overlap between engines.

…but there IS a consensus core — engines agree on the head

"No single AI ranking" is about the full top-3 list. Stratified, the engines do converge on a thin head of recognized brands:

• The same company is rank-1 on ≥2 engines in 37% of queries (≥3 in 9%, all-four in just 1%).
• If a company is top-3 on one engine, it's top-3 on a given other engine only 16% of the time — low transfer beyond the head.
• Consensus core (top-3 on all four across multiple queries): hiscox (9), the hartford (6), simply business (4), an anonymized UK accounting firm (4), mwa accounting (4), flux (3).

Concrete cases where one company was top-3 on all four engines (ranks o/a/g/p):

So the honest statement: engines share a brand-dominated head (Hiscox, The Hartford, Simply Business…) but diverge on everything below it. [source: rankmatrix.jsonl, 487 multi-engine queries]

4 · The Q&A contradiction, adjudicated

Takeaway: FAQ/Q&A formatting nudges citation up a little on every engine (+1–4%) — neither the +25% nor the −5.7% the published studies claim.

Semrush +25% vs Citation-Absorption −5.74%. Our per-LLM test:

Modestly positive on all four (+1–4%) — does not flip sign; far smaller than either camp claims.

5 · Company breadth

Takeaway: More pages looks like it helps — but only because prominent brands happen to have more pages. Control for prominence and the effect reverses. Publishing more pages is not a lever.

Raw ρ = 0.08; controlling for the company effect, partial ρ = -0.36 — apparent breadth benefit is a prominence proxy.

6 · Per-niche variation

Takeaway: The picture barely changes by niche — schema helps trades a touch more than accounting/crypto, but brand prominence dominates in every vertical.

Top-3 rate + on-page lift by niche. Variation is modest; schema helps trades (Landscapers/Carpenters/Builders ~1.08–1.09×) slightly more than accounting/crypto. The brand-dominated story holds everywhere.

7 · Per-location variation & localization

Top-3 rate by the city named in the prompt (variation tracks competitive density — bigger metros = more competitors = lower top-3 rate). Cities are resolved from the prompt text, not the stored location tag (see the data-quality note below for why).

Coverage — not every prompt names a place. 81% of insurance + accounting citations come from prompts that name a UK city (101,645); the other 19% (24,627) are national, shown as the italic National (no city) baseline. So the city rows are not the whole dataset — they're the located 81%. Schema-rich lift ≈ 1.0× in every location ⇒ on-page schema doesn't change top-3 odds anywhere. The PCB/electronics dataset is excluded entirely (no geographic prompts).

Localization — do local players beat the big national brands?

Takeaway (the hypothesis was only half right). Naming a city does not uniformly hand the top 3 to small players — across all localized prompts the national incumbents actually hold a slightly higher share than in national prompts (17% vs 12%). The real effect is city-specific: in 4 of 6 cities (Glasgow, Birmingham, Manchester and Cardiff) the incumbents nearly vanish — holding under 15% of the top 3 — while in London and Leeds they hold ~22–27%. Where local players win, they win decisively: each low-incumbent city has hundreds of distinct local/regional firms taking those slots.

The gradient is the finding: Glasgow → 0% incumbent share (local brokers own it outright), rising monotonically to Leeds 27%. The national-average comparison washes this out because London + Leeds carry most of the volume. The 15 incumbents under test: hiscox, simply business, axa, direct line for business, the hartford, axa uk, markel direct, aviva…

[source: rankings.json — de-branded, test/probe prompts removed — × query-text city × top-15 multi-city incumbent set. Insurance + accounting sites only (5,446 localized + 1,699 national top-3 slots); the PCB/electronics dataset has no geographic prompts.]

Who wins locations, per engine

National-incumbent share of the top-3, by engine × city (cities ordered most→least local-friendly). Green = local players own it; red = the established national brands hold on.

perplexity is the most open to local/niche players (incumbents only 13% of localized top-3); anthropic leans hardest on the established national brands (20%). But the gradient itself is identical on all four engines: Glasgow 0% everywhere, climbing to ~25–33% in London & Leeds.

Your thesis: is niche / regional easier to rank than global?

8 · The eligibility floor — cited vs wild

Takeaway: To be citable at all, be answer-shaped (definitions, examples — 74% of cited pages vs ~18% of the web) and carry real content schema (FAQ/Article/LocalBusiness). Boilerplate schema is everywhere and doesn't distinguish you. This earns eligibility, not rank.

% with each signal across four cohorts: cited pages · uncited siblings (same domains) · Majestic top-1M draw · a uniform Common Crawl web sample (n=106 — the true long tail Majestic excludes). Key read: the two random baselines agree, and schema is widespread (~half the live web has it — CMSs auto-emit), so schema presence is a weak discriminator. The real ~3× gap is answerability/depth (answer-shaped content, definitions, examples, length): cited 74% answerable vs CC-web 18%. So eligibility = substantive answer-shaped content + real content schema, not boilerplate markup. Note the schema decomposition below: boilerplate (WebSite/Org/Breadcrumb) is everywhere (~40–60% across all cohorts, weak signal); content schema (LocalBusiness/FAQ/Article/Review) is the ~3× differentiator (cited 46% vs web 19%–18%). Neither moves rank once you're cited.

Medians

CC web n=106 (uniform Common Crawl draw, the genuine long tail) — rates ±~10%. It tracks Majestic on schema (≈ half the web) but confirms the big answerability/depth gap, validating the eligibility-floor finding against a true-web baseline.

9 · Recency

Flat — no "3-month freshness cliff" in our data.

10 · This study vs prior work

Princeton GEO is causal (edit one page, re-query); ours is observational across a brand-dominated competitive field — both can hold. The headline divergence: authority/backlinks and stats are levers in the SEO/GEO literature but non-levers for LLM citation here.

11 · Statistical robustness — Bayesian vs frequentist cluster-robust + MCMC

Takeaway: We re-checked the findings three ways (standard, cluster-robust, and Bayesian) — they all agree. The conclusions don't depend on the statistical method; the company effect dwarfs every page feature in all of them.

schema presence

readability

Point estimates agree (Bayes ≈ frequentist at this n); cluster-robust CIs (clustered by company) are wider — the honest non-independence fix. Bayesian fixed-effects on full per-engine data; hierarchical on a 12,000-row subsample (rhat>1.01 — magnitude approximate, conclusion robust).

12 · Method & limits