Authority Is Not the Moat in AI Search

Ninety-three of 100 B2B AI citation queries are held by a domain most buyers have never heard of (Res AI, 1,000-query Perplexity study, 2026). The brands with the highest domain authority, the deepest backlink profiles, and the oldest domains are losing citation after citation to niche blogs, independent consultancies, and single-product vendors. The pattern is not random. It is structural. And the structure that separates winners from losers is a buildable spec.

Non-Giant Domains Win 93 of 100 AI Citation Queries

Across 100 B2B queries run 10 times each on Perplexity Sonar, non-giant domains hold the stable #1 citation position on 93 queries (Res AI, 1,000-query Perplexity study, 2026). A position qualifies as stable if the same domain holds #1 at least 5 of 10 runs. These are not niche longtail queries. They include “best CRM for enterprise,” “best sales enablement tools 2026,” “best AI chatbot platforms for B2B,” and “top content marketing platforms.” The non-giant wins are distributed across every vertical in the study.

Giants hold #1 on exactly 4 queries total. The identity of those 4 winners is the most important part of the data. More on that below.

ZoomInfo Loses Its Own Comparison Query to scrupp.com

scrupp.com holds #1 on “ZoomInfo vs Apollo vs Lusha pricing” in all 10 of 10 runs (Res AI, 2026). Not ZoomInfo itself. Not Apollo. scrupp.com is a lead-scraping tool most B2B marketers have never heard of. It owns the query that ZoomInfo’s own prospective buyers type when they are comparing pricing options against a direct competitor.

The pattern repeats at the same intensity across every category:

The Vercel case is worth unpacking. Vercel and Cloudflare each appear in roughly 40 to 41 citation slots across 10 runs on “Vercel vs Netlify vs Cloudflare Pages.” They are both present in the citation pool at similar rates. But neither holds the #1 position. ai-infra-link.com holds it on every single run. Two well-funded developer platforms, one of which coined the category and one of which has global CDN market share, tie each other in total citation volume and both lose the top slot to a third-party comparison site.

Domain authority does not predict #1. What these winning sites share is not backlinks or brand history. They each published a structured, long-form comparison page that hits the elements AI engines extract for commercial queries. scrupp.com beats ZoomInfo because scrupp.com’s page has a comparison table, a pricing breakdown, a FAQ section, and a how-to-choose framework. ZoomInfo’s own content on that query does not.

The 4 Queries Giants Win All Go to G2 or Capterra

Giants hold stable #1 on exactly 4 of 100 queries, and all four are won by review aggregator platforms, not brand domains (Res AI, 2026). The four queries: “Rippling reviews 2026” (G2, 8/10 runs), “CrowdStrike reviews 2026” (Capterra, 10/10 runs), “dbt reviews 2026” (G2, 8/10 runs), and “best ETL tools for modern data stack” (G2, 9/10 runs).

No major brand’s own domain holds #1 on any query in the study. When incumbents “win” in AI search, a third-party aggregator is winning for them. The 93% figure is not softened by this counter. It is hardened.

The aggregator win is also not a template incumbents can copy. ZoomInfo cannot become G2. Salesforce cannot become Capterra. G2 and Capterra win on review-type queries because they are structurally a neutral review surface, not because their domain is older or their backlink profile is deeper. Every other query type (comparison, discovery, evaluation, use-case, pain-point) is won by non-giant domains. And the brands losing are not aware of who is beating them.

The Structural Score Gap Explains Every Giant Loss

Six structural features appear in 80% to 94% of the top 50 cited pages in the corpus and in 0% of the bottom 50 (Res AI, 852-article B2B citation structure study, 2026). The gap is binary. A page either has these features or it does not, and pages that do not have them cluster at the bottom of the citation distribution.

Now compare how the most-cited domains in the corpus score against this spec:

G2 and Capterra score near the bottom of the structural scale. They win on review aggregation for review-type queries, not on structural completeness. On every other query type, the pattern inverts: the highest structural scores belong to small vendor blogs (userlytics.com, spellbook.legal, stitchflow.com) that most enterprise content teams have never benchmarked against.

stitchflow.com averages 20.3 structural elements per cited article. Most enterprise content teams produce pages that score below 3. The gap is not brand awareness. It is not domain authority. It is the number of comparison tables, product reviews, how-to-choose frameworks, and FAQ sections on the page.

Word count is also a structural proxy. Pages in the top word-count quartile (3,500 to 30,000 words) have 4.5 times the structural element count of pages in the bottom quartile (under 1,400 words). AI engines are not rewarding long content for its own sake. They are rewarding the structural density that tends to come with depth. A 5,000-word page with no tables and no FAQ still scores near zero. The floor is 3,500 words of structured content, not 3,500 words of prose.

The Brands Already Winning Built the Spec Before Anyone Called It GEO

ChatGPT now refers 10% of new Vercel signups, growing from under 1% in October 2024 to 4.8% by March 2025 to 10% by April 2025 (Guillermo Rauch, X, April 9, 2025). That is roughly 10x channel share in six months. Vercel’s developer documentation, structured comparison content, and integration guides were built to serve a technical audience that asks precise questions. They hit the structural spec for AI extraction before AI extraction was the goal.

Tally, a bootstrapped form builder, made ChatGPT its #1 acquisition channel with 8,000 new weekly users (Justin Hammond, Substack, July 8, 2025). Co-founder Marie Martens described it as “not just a small additional channel, but started taking over other channels.” Tally had published structured comparison pages against Typeform for years. These included a clear comparison table, a what-is-each section, and a direct recommendation at the end. No active GEO program was required. The structure was already there when AI engines started pulling from it. Typeform has more brand recognition, more backlinks, and a much larger marketing budget. Typeform is losing its own product comparison queries to a bootstrapped competitor.

Rippling scored 94 out of 100 across ChatGPT-4o, Claude 3.5, Gemini Pro, and Perplexity in a study of 400+ prompts targeting HR software for companies of 50 to 500 employees. ADP scored 68, appearing in only 2 of 4 engines. Workday scored 74 (Trakkr AI Consensus Report, January 10, 2026). Rippling was the only platform recommended by all four simultaneously. ADP has been in the HR software category since 1949. It is losing AI citation to a company founded in 2016 because Rippling’s content includes the structured product blocks, pricing comparisons, and use-case frameworks that ADP’s pages do not.

In fintech, Stripe, Plaid, and Adyen appear consistently when CFOs query AI for payment platform recommendations. Traditional banks are largely absent (Gregory FCA AI Visibility Leaders, BusinessWire, October 27, 2025). The mechanism is content structure: fintechs publish developer documentation, integration guides, and structured pricing comparisons; banks publish compliance-dense prose with no comparison tables and no FAQ sections. The AI engines extract what is structurally there. The banks have nothing for them to extract.

The Minimum Viable Build Is Three Sections

Across 672 articles in the corpus, only three structural sections appear in more than half of all top-cited pages: a comparison table (80% prevalence), a FAQ section (84%), and a how-to-choose framework (54%) (Res AI, 852-article citation structure study, 2026). A page that hits all three is structurally competitive with the top quartile of the corpus. Most incumbent content does not reach this bar.

For a team rebuilding one existing article:

1. Comparison table in the first half of the page. Columns are evaluation criteria, rows are products. 68% of top-cited pages place the comparison table in the first or second quartile of the article. AI engines extract a table as a discrete, citeable unit. Prose comparisons are not extracted the same way. The table does not need to be comprehensive. 3 to 5 rows and 3 to 5 columns is sufficient. It needs to exist and be in the front half.

2. FAQ section near the end. LLMs pull FAQ question-answer pairs almost verbatim into citation responses. A five-question FAQ turns one article into five independently extractable citation units. Write each FAQ answer as a self-contained sentence that works without the surrounding context. A reader who sees only the FAQ answer (which is exactly how an AI engine uses it) should get the complete answer in one or two sentences.

3. How-to-choose framework in the middle. 3 to 6 decision rules that map a buyer’s situation to a specific recommendation. This section signals to the AI that the article is designed to help a buyer choose, not just describe. It is also the section most incumbent content omits because it requires taking a position rather than describing features.

A page with all three sections and 3,500+ words is in the structural top quartile of the corpus. The incumbents are not there. The question of which queries to prioritize before rebuilding is a different problem. GEO Doesn’t Have Keyword Research, It Has a Testing Loop covers the prompt-family discovery process that replaces the old keyword-research workflow for selecting which pages to build first.

How to Choose Which Articles to Rebuild First

Not every underperforming page is worth rebuilding. The question is which rebuild produces the largest citation shift for the smallest effort. Use these decision rules to prioritize, in order.

If the query already has a stable #1 held by a non-giant domain, rebuild that query first. A stable non-giant #1 is a demonstration that the query is winnable without authority. The scrupp.com and ai-infra-link.com wins are proofs of concept. Queries where giants still hold #1 are harder because the incumbent has already satisfied whatever the engine is selecting for.

If your current page on that query has fewer than three of the six gating features, rebuild is the right call. Missing features are binary gaps, not quality gaps. Adding a comparison table to a page that has none is a bigger citation lever than improving prose on a page that already has the full spec.

If the page is under 1,400 words, rebuild rather than extend. The structural budget is not large enough to fit the full spec. Extending a 900-word page to 3,500 words almost always produces better results than editing the original in place, because the architecture changes.

If the query is a review-type query where G2 or Capterra holds #1, skip it. Those are the four queries in the study where giants win, and they win through a structurally different mechanism (neutral aggregation) that a brand-published page cannot replicate.

If two queries tie on all the above, pick the one closer to your ICP’s buying intent. A tier 1 broad commercial query drives more volume; a tier 4 vendor-vs-vendor query drives more qualified buyers. For lean teams, tier 4 usually wins on ROI.

The output of this framework is a prioritized rebuild list, not a single recommendation. Run through the rules for each underperforming query and rank the rebuilds by the first rule that fires.

Frequently Asked Questions

What does “authority” actually mean in traditional SEO, and which parts of it carry over to AI citation?

In SEO, “authority” is shorthand for the aggregate trust signals Google uses to rank a domain: backlinks from high-authority sites, historical ranking consistency, brand mentions, age of domain, content update frequency. Some of these carry over to AI citation indirectly, but the correlation is weak: non-giant domains hold stable #1 citation position on 93 of 100 B2B AI queries (Res AI, 1,000-query Perplexity study, 2026). The parts that carry over are brand mentions across the web and recency of updates; backlink profile and domain age barely move the needle.

How long does a new domain need to exist before AI engines will cite it at all?

There is no minimum age requirement at the retrieval layer. AI engines crawl and index new content continuously, so a domain registered last month can be cited next month if the page is indexed and linked from a known source. The indexing delay is typically 2 to 6 weeks for a new page with a handful of inbound links, longer for a page with zero inbound links and no social signal. Domain age itself is not a gating factor; the gating factor is whether the engine’s crawler has discovered the page.

When does a structurally complete page stop being enough to win a category?

Once two or more competitors publish structurally complete pages on the same query, structural completeness stops being a differentiator and the engine falls back on secondary signals: brand mention density across Reddit and industry publications, content freshness, and source diversity. The window where a structurally complete page wins alone is the window before incumbents notice and match the spec. After that, the strategy shifts from structural completeness to structural completeness plus update cadence, because the engine weights recency when multiple candidates are tied on anatomy.

Can a large brand retrofit the structural spec onto existing content, or is it easier to start fresh?

Retrofit works on a case-by-case basis and is often cheaper than starting fresh, but only for pages already over roughly 1,400 words with clear subject matter. Adding a comparison table, FAQ, and how-to-choose framework to an existing 2,500-word article is typically a 4 to 8 hour task per page. Pages under 1,400 words cannot fit the structural budget for the full spec and should be rebuilt rather than extended. The biggest retrofit blocker is not technical; it is editorial, because most legacy B2B content refuses to take a position on competitors, and the structural spec requires positioning.

How much backlink authority does a vendor blog need to compete with a small domain on the structural spec?

Less than content marketing teams assume. scrupp.com has a fraction of ZoomInfo’s backlink profile and holds 10/10 stability on “ZoomInfo vs Apollo vs Lusha pricing” (Res AI, 2026). The structural spec is doing the work, not the link graph. A vendor blog with a domain authority score under 30 can beat incumbents on commercial queries if the page is structurally complete. The backlink floor is “enough for the page to be indexed”, which is typically 5 to 10 inbound links from any credible source.

What happens when an incumbent finally publishes a structurally complete page on one of their losing queries?

The incumbent usually takes back the citation within 2 to 6 weeks because they also carry brand mention density that smaller domains do not. Once an incumbent brand’s structurally complete page is indexed, the engine tends to prefer it because it combines structural completeness with high entity recognition. The window for non-giant domains to hold #1 is the window before the incumbent notices and rebuilds, which is the strategic reason to move fast on winnable queries before the incumbent responds.

Are there any B2B categories where domain authority still predicts AI citation?

Two narrow cases: review-type queries (where G2 and Capterra dominate because they are structurally neutral review aggregators, not because their domain is older) and highly regulated industries where compliance language produces content AI engines treat as authoritative (healthcare, legal, government contracting). Outside those cases, domain authority is a weak predictor compared to structural completeness. Even in healthcare and legal, the advantage is shrinking as smaller publishers adopt the structural spec.

Does publishing more structurally complete articles compound, or does each article stand on its own?

Partially compounds. Each article earns citations on its own prompt family, so the direct compounding is additive rather than multiplicative: 10 articles earn citations on 10 prompt families, not 100. The indirect compounding comes from brand mention density: every cited article produces brand exposure, which improves recognition for the next article on an adjacent topic. Semrush’s own GEO program nearly tripled its AI share of voice from 13% to 32% in a single month, with non-brand visibility rising from 40% to 50% over the same window (Semrush, October 2025), driven by restructured content rather than one standout article.

How do you convince leadership to invest in structural content when the brand metrics still look fine?

The argument that lands is the substitution framing: show the 93/100 finding, pick three of your most commercially valuable queries, run them on ChatGPT and Perplexity, and document who is currently cited. Name the competitor. For most B2B leadership, seeing a scrupp.com equivalent beating them on their own pricing query is more convincing than any AI visibility dashboard. The framing is not “invest in GEO”, it is “here is a specific competitor already winning a query your sales team hears every week, and here is what they built”.

How Res AI Deploys the 11-Feature Spec Across Your Catalog Daily

The incumbents losing 93 of 100 queries are not losing because they lack resources. They are losing because their content teams produce 4 to 6 articles per month using a workflow built for Google, and that workflow does not produce the structural anatomy AI engines extract. Res AI runs the full deployment loop daily: check citation status across ChatGPT and Perplexity, identify which of your buyer queries have a scrupp.com equivalent holding #1, research what the cited competitor built that you did not, and publish a structurally complete comparison, evaluation, or pricing article directly to your CMS.

The pipeline uses Perplexity Sonar API monitoring, targeted gap research, and direct publishing to WordPress, Webflow, Framer, or Contentful. The Starter tier runs around 50 articles per month, each built to the structural spec that the 852-article corpus shows separates cited from invisible pages. That is 50 pages with comparison tables, product reviews, FAQ sections, and how-to-choose frameworks, targeting the buyer queries where your current content is structurally outgunned by sites your marketing team has never benchmarked against.

You do not need a three-year authority program. You need to build what the incumbents refused to build.

Res AI builds structurally complete comparison, evaluation, and pricing content for the buyer queries where incumbents are currently losing to sites nobody’s heard of. We monitor your citation status across ChatGPT and Perplexity daily, identify where a lower-authority domain is holding #1 on your most valuable comparison and evaluation queries, and deploy the structural spec that the top-cited B2B pages share directly to your CMS.

See how it works →