The evidence.

Every figure on the homepage, with how it was measured and what it does not show. Results that go against us are printed next to the ones that go for us. Method notes, code and raw data are available on request.

Without a source, agents miss funding facts.

We took 79 European funding rounds announced by other publications and asked two models about each one three ways: alone, with web search, and with the Funding Explorer's MCP server. Every answer was checked against the announcement.

96% 76 of 79 rounds were in the Funding Explorer (95% CI 89 to 99%), a median of 0.6 days after the announcement.

SetupCorrectWrongCost per correct answer
Model alone0%37%n/a
Model + web search97%3%$0.029
Model + Funding Explorer MCP92%3%$0.0035

79 rounds announced 13 Aug to 12 Oct 2026, from five other publications; 2 models, 158 answers per setup. Web search scored slightly higher on this sample (97% against 92%; significant for Gemini alone, p = 0.03). MCP answers cite funding.tech.eu records; 59 of 60 sampled citations state the exact figure. Two thirds of the sample comes from EU-Startups, which the Funding Explorer reads, so coverage beyond mainstream startup media is not measured. Full method and data on request.

The newsroom makes the data.

tech.eu has reported European technology since 2013. Its reporting is written into the Funding Explorer the same day, and fundingdata.com draws on 279 sources worldwide.

  • 59.7%

    of euros raised in European rounds in 2026 have a tech.eu article among their sources (36.7% of rounds).

    924 of 2,520 rounds, 1 Jan to 9 Oct 2026

  • 54%

    of rounds of €100M or more since 2023 have a tech.eu article among their sources.

    204 of 376 rounds of €100M or more

  • 86.6%

    of rounds reported by tech.eu are in the Funding Explorer on the round's date, usually the article's date (98.7% within a day).

    n = 299, rounds ingested 1 Jul to 9 Oct 2026

  • 53%

    of 2026 rounds on fundingdata.com were reported by two or more publishers; 94% cite a published article; about 11 hours from announcement to record.

    2026 rounds; announcement to record, median weighted across European countries over the last 90 days, day-precision dates

Computed 12 Oct 2026 from the open Funding Explorer API (non-debt rounds) and fundingdata.com. Each round carries one best-ranked source, so these count rounds where tech.eu is among the sources, not where it was first.

What a sourced fact is worth.

Three buyers, three costs: the agent that guesses, the analyst who searches, and the AI company that needs live, attributable data.

  • Teams building agents

    8×

    lower cost per correct answer through the Funding Explorer MCP than through web search ($0.0035 against $0.029), at 92% accuracy against 97%. Answers cite funding.tech.eu records.

    Fores Media agent test, v2, Oct 2026

  • Investors and analysts

    6+ hours

    a week spent searching for information in documents, by half of the US finance professionals surveyed. About £12,900 a year at a London associate's cost (our estimate). Incumbent data seats: Crunchbase median $20,000 a year; Dealroom €250 per seat a month.

    Hebbia survey (n = 525, US); Vendr; dealroom.co, 2026

  • AI companies

    2 → 18

    public AI content deals for live, attributable access, 2023 to 2025, with 34 projected for 2026. Reported: Reddit “other revenue”, mainly content licensing, $140M (FY2025); Wiley AI licensing $49.1M (FY2026).

    Media and the Machine; Reddit and Wiley filings, Jun 2026

How the market looks to an agent.

Once a day we run our open agent-readiness check on each homepage: can an agent read it without JavaScript, find the usage terms, get markdown, find tools to call? It measures what an agent finds without an account. Several of these companies sell API access to paying customers.

HomepageScoreReadable without JavaScriptExplicit policy for AI crawlersMachine-readable usage termsllms.txtMarkdown for agentsStructured dataTools agents can callSitemap and page metadata
Fores Mediaforesmedia.vercel.app100Readable without JavaScript: passExplicit policy for AI crawlers: passMachine-readable usage terms: passllms.txt: passMarkdown for agents: passStructured data: passTools agents can call: passSitemap and page metadata: pass
Tracxntracxn.com50Readable without JavaScript: passExplicit policy for AI crawlers: passMachine-readable usage terms: failllms.txt: failMarkdown for agents: failStructured data: partialTools agents can call: failSitemap and page metadata: pass
CB Insightswww.cbinsights.com45Readable without JavaScript: passExplicit policy for AI crawlers: partialMachine-readable usage terms: failllms.txt: failMarkdown for agents: failStructured data: failTools agents can call: partialSitemap and page metadata: pass
Crunchbasewww.crunchbase.com40Readable without JavaScript: passExplicit policy for AI crawlers: passMachine-readable usage terms: failllms.txt: failMarkdown for agents: failStructured data: failTools agents can call: failSitemap and page metadata: partial
Dealroomdealroom.co38Readable without JavaScript: passExplicit policy for AI crawlers: partialMachine-readable usage terms: failllms.txt: passMarkdown for agents: failStructured data: failTools agents can call: failSitemap and page metadata: partial

Measured 16 Oct 14:27 UTC with the check at /check, which anyone can run on any site. Each dot is one check, filled when it passes and half filled when it partly passes. Not readable by the check on this run, usually because of bot protection: Tech.eu Funding Explorer (funding.tech.eu), PitchBook (pitchbook.com), Bloomberg (www.bloomberg.com).

Research Radar, tested on the past.

Each cohort is the 100 phrases Research Radar would have flagged on that date, scored by whether their share of all research kept growing. Cohorts that failed are shown too.

Flagged inTop 100 flagged that kept growingAll phrasesMomentum top 100
2022-09scored 2022-10 to 2023-0929%95% CI 21 to 39%, all phrases 40%: not better than all phrases40%34%
2023-09scored 2023-10 to 2024-0989%95% CI 81 to 94%, all phrases 40%40%74%
2024-09scored 2024-10 to 2025-0950%95% CI 40 to 60%, all phrases 51%: not better than all phrases51%56%
2025-06scored 2025-07 to 2026-0675%95% CI 66 to 82%, all phrases 55%55%30%
2025-09scored 2025-10 to 2026-0983%95% CI 74 to 89%, all phrases 55%55%38%
2025-12scored 2026-01 to 2026-10, window still open67%95% CI 57 to 75%, all phrases 54%: provisional54%37%

Backtests computed in 2026 on past months, not a live record. Kept growing: the phrase's share of all papers rose over the scoring window. Momentum: the 100 fastest-growing phrases among those already in the top quarter by volume. n = 100 per cohort. From Research Radar's public record, as it publishes it.

Backtest examples, flagged in 2023-09: papers when flagged, then papers in the scoring window after

  • gpt-4599 → 1,995
  • segment anything182 → 493
  • prompt engineering156 → 568
  • instruction tuning164 → 737
  • low rank adaptation47 → 358
  • lk-9944 → 10miss

Selection: Picked by hand from the 2023-09 top 25 to show named methods and models; the full list also holds writing-style phrases. The highest-ranked clear miss is shown with them.

We write down our predictions before they happen.

Every Research Radar run freezes its ranked list of rising research. The list is hashed, the hash is chained to the run before it, and the chain head is stamped by two independent timestamp authorities. Each list is scored later against what the literature actually did. Nobody, including us, can rewrite an earlier list.

Chain
13 runs, every hash verified
Scores
Scored once three complete months of later data exist.
Not included
Backtests on earlier months are published separately and are not part of this record.

Check a stamp yourself

openssl ts -verify -digest <chain> -in <chain>.freetsa.tsr -CAfile freetsa-cacert.pem

FreeTSA's certificate: freetsa.org/files/cacert.pem

  1. Run 86Awaiting score

    chain5618b08f2bcd8e41d19edf21f42cdd2b9e15ca0bac476a4490fcff08168a98ff

    preve8142246de7fdd6b3e64d07dd4563ec95cc8e52f94a8d6c04e91dbfeb7272389

  2. Run 78Awaiting score

    chaine8142246de7fdd6b3e64d07dd4563ec95cc8e52f94a8d6c04e91dbfeb7272389

    prev5bee98203e2357b02881a4036437616923a04739345e06cde8c29b1d251a34da

  3. Run 70Awaiting score

    chain1446c6a77af2677926c611cd304d1f130005735da26e59844b99c43669c08cd5

    prev3580d834f3729b9f6fac03ba1d0947858666886028f904d84bb0aa8a8b66f4b7

  4. Run 64Awaiting score

    chain4e362cd93c10d82d2b08fdc540bf6d81ecd27a28ef4773584a8b83d72e383221

    prev0c2c6e0c505f80b1df03f84b1d6d126b8fb1f500a05ddfc447d0b96b91c7b3a8

  5. Run 61Awaiting score

    chain0c2c6e0c505f80b1df03f84b1d6d126b8fb1f500a05ddfc447d0b96b91c7b3a8

    prev9b282fa95506d516524c39c004684a3f07c09e7d70b1aa81f5eb52214efb093c

  6. Run 53Awaiting score

    chain9b282fa95506d516524c39c004684a3f07c09e7d70b1aa81f5eb52214efb093c

    prev605d5fdd55c9e71346fbe64f9292aea3bbd2a419f6705302b05ead20d0553cde

  7. Run 46Awaiting score

    chain605d5fdd55c9e71346fbe64f9292aea3bbd2a419f6705302b05ead20d0553cde

    prevf5efcde5f2a076615a01cb39990eea1957767bbddfe919057bd9ba6cc0537d65

  8. Run 43Awaiting score

    chainf5efcde5f2a076615a01cb39990eea1957767bbddfe919057bd9ba6cc0537d65

    preva14cba827cc3d274c99ff42af81d596877c4316ba6e6e8e5c36b865696cf373a

  9. Run 38Awaiting score

    chaina14cba827cc3d274c99ff42af81d596877c4316ba6e6e8e5c36b865696cf373a

    preva5f74bb5fc0e1c53e400de55e7f33f84303b511e3b044cc173a3bfa4006f2306

  10. Run 35Awaiting score

    chaina5f74bb5fc0e1c53e400de55e7f33f84303b511e3b044cc173a3bfa4006f2306

    prevc44a0d27b2e7f3cc0becbe0932a482373070f37c4179bffaeac3a42e6d4077e4

  11. Run 15Awaiting score

    chainc44a0d27b2e7f3cc0becbe0932a482373070f37c4179bffaeac3a42e6d4077e4

    prev0142c54f8033e3b36df1de07a49f52d7e891422e2ce780244733af1b7eec26c6

  12. Run 12Awaiting score

    chain0142c54f8033e3b36df1de07a49f52d7e891422e2ce780244733af1b7eec26c6

    prev1a22770ca14fdcc2a4fa4e10dbe3ad54e02aa19ca60a510c9b54ecd3a355a35e

  13. Run 6Awaiting score

    chain1a22770ca14fdcc2a4fa4e10dbe3ad54e02aa19ca60a510c9b54ecd3a355a35e

    prev0337ccba5a1509b8285e26ea0f99ef4cc18e3f33ca86c09163c8d89f3886a3fe