Top Publications Cited by AI Search Engines in B2B (2026)

Summary: Measured as a rate on one denominator, the publications AI search engines cite most in B2B are Medium (cited in 5.04% of 16,039 monitored answer runs) and Forbes (4.15%), both at Confidence A and both observed by all six monitored engines. TechCrunch is cited in 1.02% of runs and ranks 9 of 1,230 classified editorial publications.

Published March 30, 2026. Measured against Machine Relations Index release mri_score_v2.0+2026-09-20+47973f373a20. For the live index, see the Machine Relations Index.

AI search engines do not cite the web evenly. They cite a narrow set of surfaces repeatedly, and that pattern matters because citation concentration determines which publishers shape category understanding. What changes when you measure it properly is which surfaces those are, and by how much.

machinerelations.ai tracks this distinction because Machine Relations is about being chosen by retrieval systems, not just being crawlable. Publisher selection is one of the clearest signals in that system, and it is only useful when it is measured as a rate on a fixed denominator.

Methodology and sources #

This ranking is measured against Machine Relations Index release mri_score_v2.0+2026-09-20+47973f373a20, generated 2026-09-20 over an observation window of 2026-05-10 to 2026-09-20. The methodology is mri_score_v2.0. Every figure below is read from the public per-domain profiles at https://machinerelations.ai/index/domains/<domain>, which are the same pages any reader can open and check.

One denominator, stated once: 16,039 monitored answer runs across six answer engines (ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity), covering 22,320 observed cited domains, of which 1,230 are classified editorial publications. Google's two answer surfaces are counted as two engines because Google documents and links them as separate products — AI Overviews inside the results page and AI Mode as a distinct conversational surface, each with its own guidance for how pages appear in it. Collapsing them would understate how often a domain is reachable in Google's AI answers and overstate the breadth of a domain that only appears in one. A publication's citation rate is the share of those 16,039 runs in which the domain is cited at all. That is a rate on a fixed denominator, so it can go down as well as up, and two publications measured in the same release are directly comparable.

Three properties of this method:

1. A rate, not a count. A count rewards whoever has been observed longest. A share of a fixed run set does not.
2. Engine coverage is reported beside the rate, and below six it is a floor. A citation total that comes almost entirely from one engine's crawl measures that crawl, not what engines select. Each row below names how many of the six monitored engines observed the domain. In this release two of those six have dated collection gaps — Google AI Overviews returned nothing on 41 of the window's days and Google AI Mode on 11 — and a gap can only ever lower an engine-coverage count, never raise it. So a row reading six of six is a measurement, and every row below six is a lower bound, written here as "at least". A publication missing a Google surface in this release may have been passed over or may simply never have had the chance to be seen; this data does not separate the two, and no row below is read as evidence that an engine declined a source.
3. A confidence state, and an honest floor. Domains with too few observations to rank are marked Collecting, and they are not ranked. Reuters and TechBullion sit below that floor in this release.

A limit of the release worth naming: its source-role field classifies both medium.com and prnewswire.com as editorial publications, while businesswire.com and globenewswire.com are classified as wire and press-release distribution. This ranking is built on the measured rate and reports the release's classification as it stands rather than re-cutting the set by hand.

Correction, updated September 26, 2026. The September 25 version of this note, current until today, named two engines with dated collection gaps. There are three. Gemini recorded no citations on 22 intermittent days inside this release's window — May 11 to 15, then scattered single days from August 11 to September 16 — alongside Google AI Overviews, which recorded none from August 15 to September 24, 2026, and Google AI Mode, which recorded none from September 14 to September 24. ChatGPT, Claude and Perplexity collected throughout; one Claude date was flagged by the same scan and carries exactly one answer run, a scan artifact rather than a gap. An engine-coverage count is a count of engines observed citing a domain, so a gap can only lower it. Five of the six sub-six rows are therefore floors — The Next Web at least 5 of 6, TechCrunch at least 5 of 6, Fortune at least 4 of 6, Reuters at least 3 of 6, TechBullion at least 3 of 6 — because at least one of the three gapped engines is among the engines absent from each. One row stays exact, and is stated with the absent engine named — Business Insider 5 of 6 with Claude absent — because Claude collected every day of the window. TechCrunch and Reuters were stated as exact until today on the strength of Gemini being a continuously collected arm; it is not, and both are floors again. Nothing else changes: the cited-run counts, confidence states, editorial standings and the ranking itself are computed on the full run set and are unaffected. The citation rates are floors for a separate reason — a blind run stays in a rate's denominator while contributing nothing to its numerator — so every rate in the table understates.

The measured ranking #

How to read the editorial-class rank. The editorial-class rank above is read off the Machine Relations Index's source_role field, which in the current release combines source-type evidence with AuthorityTech's placement catalog; 96 of the 1,231 class members are admitted through the catalog, and the Index is moving editorial classification to source-type evidence alone. The class currently includes six platform domains alongside edited newsrooms: medium.com (class rank 1), amazon.com (7), prnewswire.com (8), apple.com (16), wordpress.com (239) and blogspot.com (330). At the same time Substack, Dev.to and Beehiiv — the same hosted open-publishing product as Medium — are filed as community platforms. Those six hold 1,358 of the editorial class's 13,525 cited runs, 10.04%, and they include the class's top slot. The class denominator also moves between releases (published figures carry 1,025 to 1,230) because it counts the domains classified in each release. What is unaffected: the citation rate, cited runs, days cited, engine breadth and confidence above are direct per-domain observations over a fixed run set, and are the primary measure; read the rate first.

Four findings from the measured ranking.

Medium and Forbes are the measurement, not the footnote. Medium is cited in 5.04% of runs and Forbes in 4.15%, the only two publications in this set at Confidence A, each observed by all six engines. They are cited roughly four to five times as often as everything below them. A page about which publications AI search engines cite in B2B that sets aside the two most-cited publications is not a ranking; it is a preference.

The wire does not lead, and it does not trail either. PR Newswire is cited in 1.13% of runs, against Forbes at 4.15% and Medium at 5.04% — so Forbes is cited 3.65 times as often and Medium 4.44 times as often. What is also true: PR Newswire and TechCrunch are within 11% of each other, 182 cited runs against 164. Wire distribution is not the dominant citation source and it is not a dead channel; it sits just above a top-tier tech newsroom and far below the two leaders.

TechCrunch is a five-engine source, not a six-engine one. It is cited in 1.02% of runs, rank 9 of 1,230 classified editorial publications, and it is the only publication in the top five that is not observed by all six engines — Gemini has not cited techcrunch.com in this window at all. A publication missing from an engine entirely is a different asset than one cited across all six, and a count column cannot show that.

Reuters and TechBullion sit below the evidence floor. TechBullion is cited in 12 of 16,039 runs, observed by at least three engines, and is marked Collecting. Reuters is cited in 13 runs, observed by at least three engines, Collecting. Neither has enough observation to carry a rank.

Reuters in particular is worth a sentence, because its low citation rate here is easy to misread. It is one of the most widely distributed newsrooms in the world by audience, and publisher-side research continues to find wire and agency journalism reaching readers largely through other people's surfaces rather than its own (Reuters Institute Digital News Report 2025). Being read at scale and being cited by an answer engine are different events, and a domain whose journalism reaches readers mostly as syndicated copy on other domains will under-index on a root-domain citation measure. That is a real limit of this method, stated rather than hidden.

What a sound citation ranking requires #

Any reader holding a vendor's citation leaderboard should test it against these requirements.

A count that only climbs is not a window. A genuine rolling window moves in both directions. A figure that rises monotonically under a stable window label is a cumulative total wearing a window's label, and any ranking built on it ranks measurement duration.

One crawler's coverage is not a market. When a single crawler supplies nearly all of a domain's citations, that is a statement about one index's reach, not about what six answer engines select. This is why the table above carries an engine-coverage column. The general version of this problem is now well documented from the publisher side: independent monitoring of AI bot traffic shows retrieval volume and citation volume moving independently, and publisher trade bodies have spent the past two years on exactly that asymmetry. Anything measured at the crawl is a measure of fetching, not of selection.

Being cited is not the same as being used. Controlled source-removal work presented at EMNLP 2025 shows that what an answer cites and what actually determined the answer can come apart (Zhang et al., EMNLP 2025). A publisher ranking measures the citation layer only. It is the layer buyers can influence and verify, which is why it is worth measuring, but it is not a claim about influence on the model's output.

The structural literature agrees. Large-scale analysis of AI Search Arena logs found news citations cluster among a small number of outlets, with only 9% of all citations pointing to news sources at all (Yang, 2025) — the measured table, with Reuters below the floor and two general-interest surfaces on top, fits that finding. A comparison of six LLM-based search engines and two traditional search engines found 37% of domains cited by LLM search were unique to LLM systems while credibility and selection biases persisted (Zhang et al., 2025). GEO-16 research found metadata freshness, semantic HTML and structured data were the strongest citation correlates, with cross-engine-cited pages scoring 71% higher on quality than single-engine-cited pages (Kumar and Palkhouski, 2025). Search Arena analysis found user preference is influenced by citation count even when the cited source does not fully support the claim (Miroyan et al., 2025). And work from the AI Disclosures Project found Gemini produced no clickable citation in 92% of answers in their dataset, while Perplexity often visited around ten relevant pages per query but cited three to four (Strauss et al., 2025).

That last one is the standing caveat on any publisher ranking, ours included: it ranks what engines chose to expose, not everything they consumed.

What this means for brands #

Separate syndication from authority. Here is what the measurement supports.

1. Rank on a rate, and make your vendor show the denominator. If a citation report gives you counts without the run set they came from, you cannot tell a rising number from a longer measurement. Ask what the denominator is and whether it is fixed.
2. Read the engine column before the total. A source cited by two engines and a source cited by six are different assets at the same total. TechCrunch's absence from Gemini in this window is a planning fact that no aggregate count exposes.
3. Treat the top of the distribution as narrow and unglamorous. The two most-cited publications in this set are a general business title and an open publishing platform. Machine-readable, freely retrievable, high-velocity surfaces outperform prestige on this measurement, which is a different shopping list than a traditional media plan produces.
4. Refuse the floor. Publications below Confidence are not ranked here, and a vendor who ranks them is selling you precision that does not exist.
5. Expect the measurement to update. A citation index that re-measures on each release will move as engines move. Each figure here carries its release identifier so it can be traced and compared.

For the content-side version of the same point — that self-authored proof is usually weaker than independent interpretation — see why case studies are not getting cited by AI search.

Frequently asked questions #

Which publication is cited most by AI search engines in B2B? #

Medium, cited in 5.04% of 16,039 monitored answer runs in release mri_score_v2.0+2026-09-20+47973f373a20, followed by Forbes at 4.15%. Both are observed by all six monitored engines at Confidence A. Every other publication in this set is cited in under 1.2% of runs.

Does PR Newswire lead AI citations? #

No. Measured as a rate on a fixed run set, PR Newswire is cited in 1.13% of runs against Forbes at 4.15% and Medium at 5.04%. It does sit marginally above TechCrunch, at 182 cited runs against 164.

Why rank on a rate rather than a count? #

A cumulative count rewards whichever domain has been observed longest. A rate on a fixed run set can move in both directions, so two publications measured in the same release are directly comparable.

Where does TechCrunch actually rank? #

TechCrunch is cited in 1.02% of monitored answer runs — 164 of 16,039 — and is observed by five of the six monitored engines; Gemini did not cite techcrunch.com in this window.

Why are Reuters and TechBullion unranked? #

Both fall below the index's confidence floor in this release: Reuters is cited in 13 of 16,039 runs and TechBullion in 12, each observed by at least three engines, and both are marked Collecting. A domain with that little observation cannot carry a rank.

What is the Machine Relations view of publication strategy? #

That brands need a citation portfolio rather than a single media hit, and that the portfolio should be chosen on measured citation rate and engine coverage rather than masthead prestige. Distribution surfaces create retrievability; editorial and trade outlets create machine trust. How those work together is the subject of the broader Machine Relations system and its five-layer stack.

Updated 2026-09-23: figures reflect the current Machine Relations Index methodology; see the MRI methodology and update log.