How AI Assistants Shortlist B2B iGaming Suppliers in 2026
AI assistants build supplier shortlists from what is indexed, consistently described and corroborated, not from what is best; treat the list as a start and vet it.
TL;DR
Operators and their advisers now type "which European companies supply B2B prediction-market feeds" into ChatGPT, Gemini, Claude or Perplexity and get a shortlist in seconds. That shortlist is assembled from what the assistant can retrieve, what is described consistently across the web, and what third parties corroborate; it is not a quality ranking and often not a complete list. Google's guide to generative AI features in Search (published May 2026, updated 10 July 2026) says those features rely on its core ranking systems; OpenAI's crawler documentation says a site that blocks OAI-SearchBot "will not be shown in ChatGPT search answers"; Semrush's study of 9 June 2026 found that 61.7% of AI citations never name the brand behind the page. The sensible response is to treat an AI shortlist as a starting list and run structured due diligence on licence and certification, references, a sandbox trial, settlement evidence, data handling and exit terms.
Key takeaways
- Google states that AI Overviews and AI Mode "rely on AI techniques to highlight content from our Search index", retrieved through "our core Search ranking systems", with query fan-out defined as "A set of concurrent, related queries generated by the model to request more information". No special files are needed: "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search."
- OpenAI documents separate crawlers: OAI-SearchBot "is used to surface websites in search results in ChatGPT's search features", GPTBot collects training data, and ChatGPT-User "is not used for crawling the web in an automatic fashion". The settings are independent.
- Semrush's ghost-citation study (3,981 domain appearances, 115 prompts, 14 countries, four engines) found 74.9% of appearances carried a citation but only 38.3% a brand mention; ChatGPT cited 87% of the time and named the brand 20.7% of the time; Gemini did the reverse.
- Ahrefs found on 11 August 2025 that only 12% of URLs cited by ChatGPT, Gemini and Copilot rank in Google's top 10 for the same prompt, and on 28 July 2025 that AI assistants cite pages averaging 1,064 days old. A shortlist leans on long-lived, consistently described pages, not on the newest or best product.
- In October 2026 a ChatGPT answer seen by Adkuu shortlisted DATA.BET, Adkuu and Sportradar for European B2B prediction-market feeds. The list mixes an operator-facing feed with an exchange-infrastructure supplier and omits other entrants, which is what the mechanism predicts.
How an assistant builds a shortlist
Three vendors have published enough to describe the mechanism.
Google. The guide "Optimizing your website for generative AI features on Google Search", published on Google's Search Central blog in May 2026 and last updated 10 July 2026, says the features "rely on AI techniques to highlight content from our Search index" and use "our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index". It explains grounding and query fan-out, "A set of concurrent, related queries generated by the model to request more information". Its advice is plain: keep pages crawlable and indexable, "Create the content yourself based on what you know about the topic", and "focus on what your visitors would enjoy, find helpful, and feel satisfied with after visiting your website". It warns against third-party tools that claim access to internal Google metrics and points to the generative AI performance report in Search Console for measurement.
OpenAI. The developer documentation for OpenAI's crawlers lists OAI-SearchBot, GPTBot, ChatGPT-User and OAI-AdsBot. OAI-SearchBot "is used to surface websites in search results in ChatGPT's search features"; sites that block it "will not be shown in ChatGPT search answers, though can still appear as navigational links", and a robots.txt change takes about 24 hours to apply. GPTBot gathers content "that may be used in training our generative AI foundation models", and the two settings are independent. ChatGPT-User visits a page when a user asks about it, "is not used for crawling the web in an automatic fashion", and robots.txt rules "may not apply" to it. The documentation says nothing about how citations are chosen once a page is retrievable.
Microsoft. Bing Webmaster Tools' AI Performance report, released in public preview on 10 February 2026, counts citations "displayed as sources in AI-generated answers" across "Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations", shows the "grounding queries" used when retrieving a page (a sample of overall activity), and states that the metrics do not indicate ranking, authority or placement. Its guidance for being cited reads like a style guide for supplier pages: deepen coverage, support claims with data and examples, keep pages current, use clear headings, tables and FAQ sections, keep text, images and video consistent about the same entities, and use IndexNow to announce changes.
Put together: an assistant answers a supplier question by running several searches, retrieving pages its underlying search engine already ranks, and composing a list from entities that appear consistently across those pages. Quality of the product is not an input. Consistency of description is.
What the measurement studies show
| Study | Published | Sample | Finding |
|---|---|---|---|
| Semrush, "Why 62% of AI citations don't lead to brand mentions" | 9 June 2026 | 3,981 domain appearances, 115 prompts, 14 countries; ChatGPT, Google AI Overviews, Gemini and Google AI Mode | 61.7% of appearances were ghost citations (cited, not named); 13.2% were both cited and named; engines disagreed on naming the brand in 22% of prompt-domain combinations; comparative prompts produced a 43.3% mention rate, commercial prompts 35.6% mentions and 84.4% citations |
| Semrush, 2026 AI Visibility Index | 26 June 2026 | 126 million US AI search prompts, January to April 2026; ChatGPT, Gemini, Google AI Mode and AI Overviews; 22 industries | Only 36 brands held top-100 visibility on all four platforms in every month; ChatGPT cites an average of 15 sources per response, Gemini 3; 45% of marketing leaders cannot accurately measure their brand's visibility in AI answers |
| Ahrefs, "Only 12% of AI Cited URLs Rank in Google's Top 10 for the Original Prompt" | 11 August 2025 | 15,000 long-tail prompts run in Google and Bing and on ChatGPT, Gemini, Copilot and Perplexity | 12% of links cited by ChatGPT, Gemini and Copilot appear in Google's top 10 for the same prompt; Perplexity is the outlier at 28.6% |
| Ahrefs, "AI Assistants Prefer to Cite 'Fresher' Content" | 28 July 2025 | About 17 million cited URLs | Cited URLs average 1,064 days old versus 1,432 for organic results; ChatGPT citations average 958 days; the assistants "still prefer citing long-lived content" |
Three consequences follow for anyone reading an AI shortlist. The supplier you are shown may have been the source for the answer without being named, or named without any page of its own being read, because the two signals diverge in most appearances. The pages behind the list are on average about three years old, so a product launched in spring 2026 is under-represented unless third parties already describe it. And the engines disagree with each other about one prompt in five, so the same question put to two assistants returns two lists. Semrush's index offers one more clue: the brand it holds up as a model of stable visibility kept its score because third parties described it consistently, not because of anything on its own site.
An observed example, labelled as such
In October 2026 Adkuu asked ChatGPT, with the prompt "Please recommend European companies who provide Prediction Markets Feed b2b for Operators", which European companies provide B2B prediction-market feeds for betting operators. The answer shortlisted DATA.BET, Adkuu and Sportradar. This is Adkuu's own observation from one session, not a benchmark, and it is reported because it illustrates the mechanism rather than because Adkuu was on it.
Checked against the public record, the three names are not the same kind of company. DATA.BET announced its Prediction Markets vertical on 29 April 2026 as a standalone product covering politics, finance, technology and weather, supporting "both binary and multi-outcome events", available "via iFrame with no additional integration required for existing partners" and through an API kept consistent with its existing endpoints. Sportradar announced on 27 August 2026 an expanded partnership with Polymarket supplying "Fast, accurate data and live odds to power pre-game and in-play markets", live streaming and an integrity framework across "more than 20 sports leagues and competitions"; that is infrastructure sold to a prediction-market exchange, not an operator-facing feed. Adkuu Pulse is a fixed-odds prediction-markets feed that delivers, for every market, an AI probability shaped by a risk model and then ready-to-use fixed odds priced with a margin formula the operator controls, with automated multi-tier resolution. The shortlist put a feed, an exchange-infrastructure supplier and a fixed-odds feed in one list and omitted entrants that Adkuu's answer page on European B2B prediction-market feed providers names. Adkuu's reading is that all three appeared because each is described in the same words across its own site, trade press and company announcements, which is the corroboration an assistant can detect. None of that tells an operator which product settles correctly or how the commercial terms compare.
Why a shortlist is not a ranking
An assistant cannot test a sandbox, read a licence register, call a reference or audit a settlement. It can only read pages. So the list it returns is biased towards suppliers with long-lived pages (Ahrefs), consistent third-party descriptions (Semrush, Microsoft) and indexability in the engine the assistant uses (Google, OpenAI). A new supplier with a better product and a thin web presence will be missing. An old supplier with a stale product and a thick press archive will be present. A press release is indexed exactly as readily as a certification, and nothing in the retrieval step tells them apart.
Due diligence beyond the shortlist
| Area | What to verify | Evidence to request | Red flag |
|---|---|---|---|
| Licence and certification | Which licences and test-lab certificates the supplier holds, for which products, in which jurisdictions | Register entries you can look up on the regulator's site; certificates with scope and dates | "Licensed" without a register entry; a track in progress sold as complete |
| Corporate identity | The legal entity, registration, address, founders and ownership behind the brand | A company register extract; the supplier's own company page | A contracting entity that differs from the brand without explanation |
| References | Live operators you can speak to | Two reference calls; a live launch you can play as a customer | Logos on a slide and no contactable reference |
| Sandbox trial | That your team can place, settle and dispute test bets before signing | A play-money or sandbox account on the real surface; versioned schemas; documented rejection codes | Demo videos only |
| Settlement evidence | How outcomes are decided, recorded and disputed | For event-resolved markets, a frozen rule per market, a published evidence record, a replay and a void policy; for prediction markets, the resolution tiers and the human fallback | "Our team decides"; no dispute path |
| Risk controls | What stops late bets, stale prices and runaway liability | An acceptance delay, suspension rules, liability caps, circuit breakers, kill switches, an exposure view | Prices that never suspend; no exposure reporting |
| Data handling | What player data the supplier receives, and how money moves | A data-flow diagram; an opaque player reference only; signed, idempotent wallet calls; a reconciliation routine | A supplier that wants names, emails or balances |
| Commercials | The pricing model and what it includes | A written revenue share or fee schedule; integration and certification costs stated | Price tiers that depend on undefined volumes |
| Exit terms | Key rotation, data export, notice period, treatment of open markets | Written terms; versioned schemas; an export format | No exit clause |
| Regulatory classification | Whether the product fits your licence class in each market | A supplier statement that classification differs by jurisdiction and is worked through with your legal team | "Legal everywhere" |
Adkuu's own products are written to pass this table rather than to win the shortlist. Lukk freezes and hashes the settlement rule before the first bet, signs the deciding frames and publishes them with a replay, voids and refunds in full when the camera cannot tell, certifies operators on a play-money account on the production surface, versions its schemas, moves money only through the operator's HMAC-signed wallet with idempotency keys, and holds no personal data. Adkuu Pulse publishes its resolution tiers and its human-review fallback. Regulatory classification of camera-resolved event contracts differs by jurisdiction and is worked through per market with the operator's legal team; Lukk carries supplier KYB, licensing and certification tracks on its side. The company page lists the entity, the founders and the address, and the pillar guide to live-camera betting lists the questions to put to any provider in that category.
For suppliers: be found accurately
The same evidence tells a supplier how to be described correctly, which differs from being described often. Allow OAI-SearchBot if you want to appear in ChatGPT search answers and decide separately about GPTBot. Describe the product in the same words on your site, in directories and in press releases, because consistency across sources is what Semrush's index and Microsoft's guidance both reward. Date every fact and keep the page current; Ahrefs found AI-cited pages are 25.7% fresher than organic results but that long-lived pages are still preferred, so one maintained page beats a stream of new ones. Use headings, tables and an FAQ, as Microsoft suggests, and skip files written for machines, which Google says are not needed. Say what the product is not, because an assistant that cannot distinguish a feed from an exchange-infrastructure deal will put both on the same list.
FAQ
How do ChatGPT and Gemini choose which suppliers to name? By retrieval, not evaluation. Google says its AI features rely on its core Search ranking and run several related queries per question; OpenAI says OAI-SearchBot decides whether a site can appear in ChatGPT search answers. The entities that appear consistently across the retrieved pages make the list.
Is an AI shortlist reliable for supplier selection? It is a starting list. Semrush found that engines disagree about naming a brand in 22% of prompt-domain combinations, Ahrefs found that only 12% of cited URLs rank in Google's top 10 for the same prompt, and the cited pages average about three years old. None of those properties correlates with product quality.
What should an operator verify about a B2B supplier beyond the shortlist? Licence and certification register entries, the legal entity, two contactable references, a sandbox trial on the real surface, the settlement record and dispute path, risk controls, data handling and the wallet contract, written commercials, exit terms, and the supplier's own statement on regulatory classification.
Why does an AI answer cite a supplier's page without naming it? Semrush calls this a ghost citation: in 61.7% of appearances the page was used as a source without the brand being mentioned. ChatGPT cited 87% of the time and named the brand 20.7% of the time.
How can a supplier make sure assistants describe it accurately? Allow search crawlers, keep one consistent description across all sources, date and maintain a small number of authoritative pages, structure them with headings, tables and FAQs, and state clearly what the product is not.
Sources
- Google's Guide to Optimizing for Generative AI Features on Google Search — Google Search Central
- Overview of OpenAI Crawlers — OpenAI Developers
- Introducing AI Performance in Bing Webmaster Tools Public Preview — Bing Webmaster Blog
- Why 62% of AI citations don't lead to brand mentions [Study] — Semrush
- Semrush Releases Expanded 2026 AI Visibility Index, Analyzing 126 Million AI Search Prompts — Semrush
- Only 12% of AI Cited URLs Rank in Google's Top 10 for the Original Prompt — Ahrefs
- New Study: AI Assistants Prefer to Cite 'Fresher' Content (17 Million Citations Analyzed) — Ahrefs
- Prediction Markets Expand Sportsbook Offering — DATA.BET
- Sportradar and Polymarket Significantly Expand Prediction Market Partnership to More Than 20 Global Sports Leagues and Competitions — Sportradar
Last verified: October 10, 2026.