ENTITY SEARCH
Check AI visibility Score your brand
Contents

Lightpaper · Entity search practice

Be found.
Be cited.
Be chosen.

Entity search is the practice of making a brand unambiguous, corroborated and machine-retrievable, so search engines can find it, AI answers can cite it, and the agents now researching and buying on people’s behalf can choose it.

Read the lightpaper Score your brand Check AI visibility
  • GEO and AI search
  • Agentic and swarm search
  • Entity authority

By , Principal Consultant at Stella Rising. October 2026. About an 18-minute read.

Executive summary

Search now has three kinds of reader. Each decides differently, and all three start by working out which entity you are.

People still scan results. AI answer engines now read on their behalf and write a single answer. And increasingly, agents research, compare and act for the buyer, sometimes as dozens of sub-agents working in parallel. None of these readers can recommend a brand they can’t identify, and none will repeat claims that nothing else on the web supports.

This paper sets out a five-component model for entity search, the evidence behind each component, what changes when the searcher is an agent, a seven-step implementation playbook, and two tools: a diagnostic for finding the component that’s holding a brand back, and a checker that shows what AI crawlers can actually read.

  1. 01

    The reader changed

    Answer engines and agents now sit between a brand and its buyers. They read more sources than any person, and they cross-check.

  2. 02

    The unit changed

    Pages get ranked and cited. Entities get mentioned and chosen. Visibility now depends on what the whole web says about you, not just on your best page.

  3. 03

    The model

    Five components and two outcomes: being mentioned or chosen as a brand, and having your own pages cited. Within each outcome, the weakest component tends to set the ceiling.

  4. 04

    The measure

    AI answers vary from run to run, so track how often you appear across many prompts and runs, not a position.

What changed in 2026

When I wrote the first entity SEO research model, AI answers were still a layer sitting on top of search. Here’s how far things moved in 2026 alone.

Jan 2026 · Protocol launch

Live

Agentic checkout arrived in Google.

Google’s Universal Commerce Protocol launched in January with Shopify and retail partners. By mid-year it powered buying directly inside AI Mode and the Gemini app.

Botify, July 2026

Q1 2026 · Starting point

<2%

AI was a rounding error for retailers.

Online retailers interviewed by Digital Commerce 360 said less than 2% of their referral traffic came from AI at the start of the year. The Gemini app had about 750 million monthly users.

Digital Commerce 360; Alphabet Q4 2025

May 2026 · Under a year after launch

1B+

AI Mode passed a billion monthly users.

Google reported the milestone about a year after AI Mode’s launch, alongside more than 2.5 billion monthly users for AI Overviews.

Alphabet Q2 2026; Google I/O 2026, via OnMSFT

Jun 2026 · Forecast was 2027

57.5%

Bots passed human web traffic.

Automated requests overtook human traffic on Cloudflare’s network for the first time. Cloudflare’s CEO had expected the crossover at the end of 2027, then early 2027.

Cloudflare Radar, via WorkOS

Jul 2026 · Up from ~750M

950M

Gemini scaled, and Google merged its AI search.

Alphabet reported 950 million monthly users for the Gemini app and said AI Overviews and AI Mode now run as one search experience.

Alphabet Q2 2026 earnings

Aug 2026 · In one month

+27%

AI agent traffic hit a new high.

HUMAN’s monthly tracker recorded 27% growth in agentic traffic in August alone, with OpenAI products accounting for roughly 31% of it.

HUMAN Security, September 2026

Sep 2026 · Up from under 2%

~5%

AI’s share of retail referrals more than doubled.

Retailers interviewed by Digital Commerce 360 now see about 5% of their referral traffic coming from AI, and one expects 10% by early 2027. Referrals are a subset of all traffic, so this is larger than AI’s share of total sessions.

Digital Commerce 360, September 2026

What entity search is

Moving from strings to things, on your site and across the web.

An entity is a distinct thing: a brand, product, person, place or concept. Search and AI systems treat entities as units of meaning, not as strings of keywords. Entity search makes sure every page, profile and mention points unambiguously to one canonical entity, and that the relationships between entities are explicit: this product, from this brand, in this category, made with this ingredient, recommended by this expert.

On your own site, that means structured data with stable identifiers, one primary entity per page, and internal links that describe relationships. The result is a small knowledge graph where every page is a node. Off your site, it means the rest of the web describing you the same way. That second half is where most brands now fall short, and where AI systems look hardest.

Keyword-era SEOEntity search
Unit of optimizationTarget keywords and individual pagesCanonical entities and the relationships between them
How relevance is judgedTerm matching, links and ranking signalsClear identity, plus consistent facts corroborated across sources
What success looks likeRankings, sessions and clicksShare of AI answers, citations and agent shortlists, plus the clicks that still matter
Site architectureTopical folders and categoriesA mini knowledge graph: every page a node with a stable @id
Main failure modeAmbiguity and keyword stuffingContradictory facts across sources, and content machines can’t retrieve

Three readers, one entity

How a single question travels through each kind of search, and where your facts get tested.

  1. 01

    Search engines

    Rank pages for a query and hand the person a list. One strong page can win. Still the largest source of referral traffic by a wide margin.

  2. 02

    Answer engines

    Split the question into sub-questions, retrieve passages for each, and write one answer with a few citations. Many sources contribute; few get clicked.

  3. 03

    Agents and swarms

    Plan the task, send sub-agents to research parts of it in parallel, reconcile what they find, then shortlist, compare or buy.

Scroll sideways to see sources and outcome

Diagram of how one buyer question reaches sources in each kind of search

Answer engine. The question fans out into sub-questions, each answered from whichever source has the clearest passage. Your site is one input among many, and the answer can cite a review site or a video instead of you.

Illustrative. Real fan-outs vary by engine, query and run.

The numbers behind the shift

  • 2.5B+

    Monthly users of Google’s AI Overviews, per figures Google shared at I/O 2026.

    Google I/O 2026, via OnMSFT

  • 8% vs 15%

    Visits that ended in a click on a traditional result when an AI summary appeared, versus when none did.

    Pew Research Center, 2025

  • −58%

    Average click-through rate for the top-ranking page when an AI Overview is present.

    Ahrefs, December 2025

  • 37.9%

    Share of URLs cited in AI Overviews that also rank in the top 10, down from roughly three in four in an earlier study.

    Ahrefs, 2026

  • 7,851%

    Year-over-year growth in traffic from AI agents and agentic browsers, from a small base.

    HUMAN Security, 2026

  • +42%

    How much better AI-referred shoppers converted than other traffic on U.S. retail sites in March 2026.

    Adobe Analytics, April 2026

Two counterweights. Classic search remains the main traffic pipe: Ahrefs estimates Google sends websites roughly 190 times more traffic than ChatGPT [Ahrefs]. And AI traffic is still small next to organic search, even where it converts well. Entity search builds on SEO. It doesn’t replace it, and Google says the same: its guidance describes optimizing for AI Overviews and AI Mode as still SEO, because those features are grounded in its core ranking systems [Google]. The twist is that they rank pages for the fan-out sub-questions, not just your head term.

Agentic and swarm search

When many machines check your claims at once, and some of them can buy.

Fan-out is already standard. Google describes AI Mode as breaking a question into subtopics and issuing many searches at once, and its Deep Search mode as issuing hundreds [Google]. Research agents go further. In Anthropic’s published design, a lead agent hands sub-questions to sub-agents that search in parallel; that setup beat a single agent by 90.2% on Anthropic’s internal research evaluation [Anthropic]. Moonshot’s Kimi K2.5 can direct up to 100 sub-agents across as many as 1,500 tool calls [Moonshot AI]. This is swarm search: breadth-first research, run in parallel, then reconciled.

What changes for brands

  1. 01

    Every query becomes long-tail

    A buyer types one question; the system asks dozens of narrow ones about price, compatibility, safety, reviews and alternatives. A brand has to win many small retrievals, not one head term.

  2. 02

    Agreement gets checked

    Sub-agents gather facts independently and a synthesizer reconciles them. In my view, a spec that differs between your site, a retailer and a review looks like uncertainty, and uncertain options are the easiest to drop from a shortlist.

  3. 03

    Agents act, so data must be transactable

    Google’s Universal Commerce Protocol, launched in January 2026 with Shopify and retail partners, powers buying inside AI Mode and Gemini. OpenAI’s Agentic Commerce Protocol, built with Stripe, handles checkout in ChatGPT [Botify]. Structured product data is how these agents see a catalog.

  4. 04

    Swarms likely go where decisions are expensive

    Multi-agent research uses about 15 times the tokens of a chat [Anthropic], so my read is that it’s spent on decisions worth the cost: vendor shortlists, high-ticket purchases, health and finance research. Those are the categories where entity gaps cost the most.

The five-component model

Each component answers one question a machine must resolve before it reuses your information. They drive two different outcomes, and within each one the weakest component tends to cap the rest.

Outcome A

Mentioned and chosen

The brand is named in answers and picked by agents. Driven by integrity, corroboration and distribution. Strong brands can win here even with weak sites, because answers draw on what others publish and what models already learned.

Outcome B

Cited

Your own pages are the source an answer links to. Gated by coverage and retrievability: if a crawler can’t fetch the answer, or you don’t have it, someone else’s page gets the citation.

The ceiling idea is a planning heuristic, not a law. Components interact, and a very strong one can partly offset a weak one. But in practice the weakest link inside each outcome is where effort pays back fastest.

Component 01

Entity integrity

Can a machine tell exactly who you are, and trust what you say about yourself?

Why it matters

An engine has to resolve a name to one entity before it can recommend it. Ambiguity means omission, or being merged with a namesake. The newer risk is narrative capture: in Ahrefs’ testing, most AI tools repeated false brand claims that had been planted online [Ahrefs]. An undefined entity gets defined by whoever publishes about it.

What to do

  • Write one canonical description, fact set and naming convention, and use it everywhere: site, social bios, retailer storefronts, press kit.
  • Ship Organization and Product schema with stable @id values and sameAs links to every profile you control, plus Wikidata where you qualify.
  • Give proprietary products, methods and frameworks their own pages and identifiers, so they exist as entities rather than phrases.
  • Make experts real entities: named authors with credentials and consistent profiles.
  • Ask the major engines monthly what they believe about you, and correct wrong facts at the sources they cite.

For agents: stable identifiers (schema @id, GTINs, SKUs, merchant IDs) let parallel sub-agents confirm they’re all looking at the same product.

Failure mode: systems can’t safely reuse information they can’t attribute.

Component 02

Answer coverage

Do you own a clear, citable answer to the real question space, including the sub-questions machines generate?

Why it matters

AI search retrieves passages, not pages. Length doesn’t win: Ahrefs found near-zero correlation between content length and AI Overview citations, with more than half of cited pages under 1,000 words [Ahrefs]. Volume doesn’t win either: the number of pages on a site barely correlates with AI brand visibility [Ahrefs]. What wins is the specific answer, stated plainly, where a system can lift it.

What to do

  • Research prompts as well as keywords, and map the follow-up questions that sit behind each one.
  • Build canonical source-of-truth pages for specs, pricing, comparisons, risks and edge cases. Own your comparison pages before a competitor or affiliate does.
  • Write answer-first passages that make sense out of context, so a single chunk can stand as a citation.
  • Keep one primary entity per page, aligned across title, H1 and mainEntityOfPage.
  • Refresh on a schedule. Stale facts are the easiest ones to replace with a third party’s.

For agents: answer the constraint questions agents decompose into: availability, compatibility, sizing, shipping, returns and policy. These are rarely on marketing pages.

Failure mode: every gap you leave is filled by someone else’s page.

Component 03

Corroboration

Do credible, independent sources say the same things about you?

0.664

Correlation between branded web mentions and brand visibility in AI Overviews, the strongest factor in a study of 75,000 brands. Brands in the top quartile for mentions averaged 169 AI Overview mentions; the next quartile averaged 14.

Ahrefs, 2025

~0.737

Correlation between YouTube mentions and AI visibility, the highest of any factor across ChatGPT, AI Mode and AI Overviews.

Ahrefs, December 2025

Why it matters

Corroboration is what independent sources say about you; distribution, next, is where you place your own facts. Authority is inferred from agreement, not claimed. Mentions count whether or not they’re linked, and often they aren’t: when Ahrefs itself appears in AI answers, it gets a link only about 28% of the time. For top-of-funnel questions in ChatGPT, “best of” list articles made up 43.8% of sources [Ahrefs]. These are correlations, not proof of cause. Bigger brands have more mentions and more AI visibility for the same underlying reason, so read these numbers as a picture of what consensus looks like, not as a promise that mentions alone will move you.

What to do

  • Pull the domains engines cite for your category’s prompts. That list is your PR target list.
  • Run digital PR to get named, not just linked: lists, reviews, roundups and expert commentary.
  • Earn creator coverage, so independent videos name the brand out loud and in titles and transcripts.
  • Keep claims consistent so corroborating sources agree with each other and with you.

For agents: parallel sub-agents each find their own sources. The more independent sources agree, the more likely every sub-agent comes back with the same answer about you.

Failure mode: claims nobody else repeats read as marketing, not fact.

Component 04

Surface distribution

Does your knowledge exist on the surfaces where answers are assembled and decisions are made?

13.7%

URL overlap between citations in Google’s AI Mode and AI Overviews, even though the two reach the same conclusion 86% of the time. Engines converge on a consensus answer but draw it from different places.

Ahrefs, 730,000 response pairs

Why it matters

No single page or domain captures your share of the answer. Google’s AI surfaces cite YouTube more than any other domain [Ahrefs], and other engines lean on lists, reviews, forums and retailer pages. A coordinated content cloud beats isolated excellence.

What to do

  • Map citations by engine for your category. ChatGPT, AI Mode and AI Overviews will each need a different source mix.
  • Treat your own video as text: chapters, accurate transcripts, and the brand and product named in the words spoken.
  • Give retailers the same canonical facts you publish, so retail product pages corroborate rather than contradict you.
  • Plan depth by surface: short answers on social, longer explainers on video, canonical detail on your site.

For agents: product feeds and commerce protocols are surfaces too. Merchant Center attributes and protocol readiness decide whether a shopping agent can see and select you at all.

Failure mode: you’re the best answer on a surface no engine reads.

Component 05

Retrievability and readability

Can crawlers and agents reach, fetch, parse and act on your content without guessing?

0

OpenAI and Anthropic crawlers seen running JavaScript in Vercel’s analysis. GPTBot and ClaudeBot fetched script files but never executed them, so content that only appears client-side was invisible to them. Googlebot, which feeds Google’s AI features, and Applebot do render JavaScript.

Vercel and MERJ, December 2024

Why it matters

This is the gate in front of the citation outcome. Fetchable comes first; parseable comes second. On schema, the evidence is mixed. Microsoft has said schema markup helps its LLMs understand content [Search Engine Land]. But when Ahrefs tracked 1,885 already-cited pages that added JSON-LD, citations didn’t rise in AI Mode or ChatGPT and dipped slightly in AI Overviews [Ahrefs, via SEJ]. Treat schema as entity resolution and product eligibility, not a citation lever, and never let a fact exist only in markup.

What to do

  • Audit robots.txt and CDN or firewall bot rules for GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Google-Extended. Decide deliberately between training and search crawlers.
  • Server-render core content. Fetch key pages with JavaScript off and confirm prices, specs, FAQs and reviews are in the HTML.
  • Keep structured data consistent with what’s visible on the page; AI systems read the visible text when they fetch.
  • Publish transcripts, alt text and consistent terminology so there’s nothing to infer.

For agents: fetchers need server-rendered HTML; browser agents need a stable page structure, labeled forms and predictable checkout flows. Emerging standards such as WebMCP, still in early preview, let a site declare actions agents can take [Botify].

Failure mode: machine ambiguity caps visibility regardless of content quality.

Run the AI visibility checker

Implementation playbook

Seven steps, in order. Each one builds on the last.

  1. Step 01

    Build the entity map

    List the core entities: brand, product lines, products, ingredients or components, people, concepts. Record public identifiers such as Wikidata IDs, assign internal IDs to proprietary items, and document the relationships between them. This map is the source of truth for pages, links and schema.

  2. Step 02

    Give every page one primary entity

    Align the title, H1 and mainEntityOfPage on the same entity. Pages that try to be about several things dilute the signal for all of them.

  3. Step 03

    Connect the graph

    Use schema with stable @id values, sameAs, about and mentions, and internal links with descriptive anchors. When a related entity could be ambiguous, give enough context to disambiguate it.

  4. Step 04

    Map the fan-out

    Build a prompt set that mirrors how your category is actually asked about, list the sub-questions behind each prompt, and close the gaps with answer-first passages and source-of-truth pages.

  5. Step 05

    Earn corroboration

    Map the domains engines cite for those prompts, then earn named mentions there through PR, reviews, expert commentary and creator partnerships.

  6. Step 06

    Open the gates

    Fix crawler access and rendering, align retailer and feed data with your canonical facts, and assess readiness for the commerce protocols relevant to your category.

  7. Step 07

    Measure and repeat

    Track share of answers, citations and accuracy across engines, segment AI and agent traffic, and send each gap back to the component that explains it.

Measuring entity search

Track share of answers, not position.

Rank tracking assumes a stable list. AI answers aren’t stable. In a study of roughly 3,000 prompt runs, ChatGPT and Google’s AI returned the same list of brands less than once in a hundred times, and the same list in the same order closer to once in a thousand. What did hold steady was how often each brand appeared across runs [SparkToro].

Google now helps with part of this. Search Console’s generative AI performance reports, available to all sites since August 2026, show how often your URLs appear in AI Overviews and AI Mode [Google]. That covers citations on Google; it doesn’t show mentions, other engines, or what the answers say.

So the measurement loop runs each prompt many times per engine and reports four things: how often you’re mentioned, how often you’re cited, whether what’s said is accurate, and which component explains each gap. Fix the lowest component first, then measure again.

Pair that with outcomes. AI referrals are growing fast from a small base: Contentsquare saw them rise 632% year over year while organic search traffic fell 9%, yet they were still only about 0.2% of all sessions across the sites it measured [Contentsquare]. Report them as a quality channel, by conversion and revenue per visit, alongside branded search demand and agent activity from server logs.

  • Share of answersPercentage of runs where the brand appears, by engine and prompt cluster
  • Citation sharePercentage of answers that link to a brand-owned source
  • AccuracyPercentage of brand mentions with correct, current facts
  • AI referral valueConversion rate and revenue per visit from AI sources
  • Agent activityAgent and crawler requests from logs, by page type and outcome

Score your brand

Rate each component from 0 (absent) to 5 (best in category). You get a ceiling for each outcome: being mentioned and chosen, and being cited. The dashed line marks your lowest component.

Move a slider, or start from an example profile.

Component scores on a pentagon, with the ceiling shown as a dashed line

Average 3.0Mentioned 3Cited 3

All five are level. Raise them together, or start with the one your measurement flags most often.

Example profiles are illustrative. The score is a planning aid, not a prediction of visibility.

AI visibility checker

See what AI crawlers can and can’t read, on this page or any other. Outside Google and Apple, the main AI crawlers don’t run JavaScript, so what they get can differ sharply from what people see.

Any URL

Check a page the way AI crawlers fetch it

Fetches the page as GPTBot, ClaudeBot and a regular browser without running JavaScript, reads the site’s robots.txt rules for ten AI and search crawlers, and checks the signals machines rely on.

How the score works
  • Crawler gets the page, 25 points. A clean 200 response to a request identifying as GPTBot. If only the bot request is refused, it scores 10 and is marked unverified, because firewalls can tell this checker from the real crawler.
  • Search bots allowed, 20 points. Share of the five search-facing crawlers (OAI-SearchBot, Claude-SearchBot, PerplexityBot, Googlebot, Bingbot) that robots.txt allows for this URL.
  • Text in raw HTML, 30 points. Full marks at 400 or more words without JavaScript; capped at 5 if the page looks like an empty app shell.
  • Machine-readable signals, 25 points. Title, meta description, H1 and a same-site canonical (5 each), plus JSON-LD (5). A noindex costs 10.

The score is indicative, built from published evidence about what crawlers need. It is not a ranking factor and doesn’t predict citations.

This page

Run it on this lightpaper

Compares what you see with the raw HTML a crawler fetches, scores the overlap, and highlights anything that only appears after JavaScript runs. Here, the only flags should be labels inside the interactive diagrams. For comparison, Adobe’s own checker put the average U.S. retail homepage at 75% [Adobe].

Any site you open

Take it with you

Drag this button to your bookmarks bar, open any page, then click the bookmark. You get the same comparison and highlights on that page.

AI visibility check

Desktop browsers. On sites with strict security policies a bookmarklet may be blocked; use the URL check instead.

Sources

Figures current as of October 2026. Secondary reports are marked. Correlation studies show association, not cause.

  1. Q2 2026 earnings call: Remarks from our CEO. Google, July 2026.
  2. Google I/O 2026 usage figures. OnMSFT, May 2026. Secondary report of Google’s keynote.
  3. Alphabet earnings, Q4 2025: CEO’s remarks. Google, February 2026.
  4. Introducing Search generative AI performance reports in Search Console. Google Search Central, June 2026.
  5. Optimizing your website for generative AI features on Google Search. Google Search Central, 2026.
  6. How AI referrals impact traffic at 4 online retailers. Digital Commerce 360, September 2026.
  7. AI in Search: Going beyond information to intelligence. Google, May 2025.
  8. Google users are less likely to click on links when an AI summary appears in the results. Pew Research Center, July 2025.
  9. Update: AI Overviews reduce clicks by 58%. Ahrefs, December 2025.
  10. An analysis of AI Overview brand visibility factors (75K brands studied). Ahrefs, 2025.
  11. Top brand visibility factors in ChatGPT, AI Mode, and AI Overviews. Ahrefs, December 2025.
  12. AI Search Benchmark Report, Q1 2026. Ahrefs, 2026. Covers 13 studies.
  13. Are AI Mode and AI Overviews just different versions of the same answer? Ahrefs, 2026.
  14. Update: 38% of AI Overview citations pull from the top 10. Ahrefs, 2026.
  15. Schema markup didn’t move AI citations in Ahrefs test. Search Engine Journal, May 2026. Report of Ahrefs’ controlled study of 1,885 pages.
  16. 2026 State of AI Traffic and Cyberthreat Benchmark Report. HUMAN Security, March 2026.
  17. State of Agentic Traffic, August 2026. HUMAN Security, September 2026.
  18. AI agents now make up the majority of web traffic. WorkOS, June 2026. Secondary report of Cloudflare Radar data.
  19. How we built our multi-agent research system. Anthropic, 2025.
  20. Kimi K2.5: Visual agentic intelligence. Moonshot AI, 2026.
  21. What is Google’s Universal Commerce Protocol (UCP)? Botify, July 2026.
  22. The rise of the AI crawler. Vercel and MERJ, December 2024.
  23. Microsoft Bing/Copilot use schema for its LLMs. Search Engine Land, March 2025.
  24. AI recommendation lists repeat less than 1% of the time. Search Engine Land, 2026. Secondary report of SparkToro and Gumshoe research.
  25. AI referrals surge as web traffic falls and costs rise. eCommerceNews, February 2026. Secondary report of the Contentsquare 2026 Digital Experience Benchmark.
  26. AI traffic surge: retail sites not machine-readable. Adobe, April 2026. Based on Adobe Analytics data covering over a trillion visits to U.S. retail sites.