Methodology

How the audit was done

The population, the census design, the rubric, the decision tree, the verification protocol, the quality control, and the correction policy. Enough to run it again.

Fall 2026 edition. Bureau listings captured October 1, 2026. Next edition: Spring 2027.

Step one

Population and sample

Population. Individuals publicly listed under an artificial intelligence category page on the websites of major speaker bureaus serving U.S. corporate and association buyers, as those pages stood during the collection window. A bureau was included only after confirming it had a public AI category page.

Why the bureaus are not named. This is not a bureau comparison and no bureau-level result is published, so naming them would invite a reading the data cannot support. How many of the included bureaus list each person is recorded, and reported only in aggregate.

Bureaus included: 6
Unique people after de-duplicating: 423
Coded: 423 all of them
Design: Census
Exclusions: 0

Frozen October 2, 2026. Every figure on these pages is generated from that frozen export, not transcribed.

Every individual listed on each included bureau’s AI category page was captured on October 1, 2026, and each page was archived with its capture date.
Listings were de-duplicated across bureaus, recording for each person how many included bureaus list them.
The frame came to 423 people, and every one of them was coded. No sample was drawn, so there is no seed, no stratification and no sampling error to report.
The full text of each listing was re-read the day after capture, to judge superlatives, fee publication and whether AI is the primary topic. One listing had disappeared by then, so those three measures run on 422 rather than 423. Each figure carries its own base.
No individual was excluded. No individual was excluded. Ryan Vet appears on none of the AI category pages captured, so no exclusion was needed on his account either.

One record, coded end to end

Rubrics are easy to publish and hard to trust. This is a single record worked all the way through, so the judgment is visible rather than asserted. It is Ryan Vet, the author, and the reason is in the next line.

The author is not in the study: he appears on none of the AI category pages captured. His record is used here so that no real subject has a coding decision published about them, and because it fails one of the two tests, which is the more instructive outcome.

Hands-on test

Pass

Requires that the person built, deployed, managed the deployment of, or professionally taught hands-on use of AI, evidenced by a dated third-party source. A bio does not pass.

What carried it

Kingscrowd, June 1, 2019, describes the matching system at his company Boon as "a sophisticated AI algorithm". Third party, dated, specific.
PR Newswire launch release, June 3, 2019, describes the matching algorithm and credential verification at launch.

What did not count

His own LinkedIn entries describing machine learning at two earlier companies are self-reported, so they were read as supporting detail and not as the evidence that carries the verdict.

Study test

Fail

Requires a research appointment with three or more peer-reviewed AI publications, or at least 24 months of dated, sourced AI-focused analysis. The 24-month branch was tightened during coding so that opinion columns do not pass.

What carried it

A weekly essay series on AI and generational change, dated and sourced, running since at least September 2025.

What did not count

Under the tightened rule that series reads as analysis and opinion rather than primary research, so it does not pass. No peer-reviewed AI publications were found. A university course beginning in 2027 post-dates the capture and was not counted.

Classified as: Practitioner
AI work before Nov. 30, 2022: Yes, on the 2019 Boon evidence

Practitioner, not Practitioner-Researcher, because one of the two tests fails. That is the rubric working: it is designed to be unflattering where the evidence is thin, including to the person who wrote it.

Inclusion rules

Every individual in the AI category is included, whether AI is their primary topic or one of several. Which applies is recorded. This is deliberate: it reflects what a buyer browsing the category actually sees.
Organizations, panels, and entertainment acts are excluded.
Pre-generative machine learning, data science, predictive analytics, and robotics work all count as AI-related work. The study is not about generative AI specifically.
Speakers who represent themselves are not in this study. A separate, smaller sample of them is under consideration for a later edition.

Step two

The classification rubric

Each speaker is classified by the source of their AI knowledge, on publicly verifiable evidence only. The labels describe evidence, not quality, and a speaker in any category can be the right speaker for an event.

The five classifications.
Classification	Definition	Evidence that satisfies it
Practitioner	Verifiable hands-on AI work: built, deployed, managed the deployment of, or professionally taught the hands-on use of AI systems, with accountability for results.	Product or company pages, press coverage of a deployment, patents, code repositories, dated role history, dated course listings.
Researcher	Verifiable disciplined study of AI without hands-on work: academic research, industry analysis, or sustained AI-focused journalism.	A faculty or lab appointment, peer-reviewed AI publications, analyst reports, or a dated body of sourced AI reporting.
Practitioner-Researcher	Meets both the hands-on test and the study test.	Evidence from both of the rows above.
Secondhand Explainer	Neither test is met in public. The AI knowledge that can be verified appears to come from others’ work, courses, or the tools themselves.	Only self-description, general commentary, or content with no verifiable hands-on or research history behind it.
Insufficient evidence	Too little public information to classify with confidence. Reported as its own category and never merged into another.	A one-line listing, no site, and nothing findable.

The decision tree

Is there enough public information?

If not, the record is classified as Insufficient evidence and coding stops there.
The hands-on test

At least one verifiable role or project in which the person built, deployed, managed the deployment of, or professionally taught hands-on use of AI. Verification requires a third-party or dated source. A bio alone does not pass.
The study test

A research appointment with at least three peer-reviewed AI-related publications, or at least 24 months of dated, sourced AI-focused analysis, research, or journalism.

Amended Tightened Sept. 17, 2026. The 24-month branch passes only on peer-reviewed research, primary research, original sourced reporting, or analyst reports. Opinion columns and company marketing blogs do not pass. Every record coded before the change kept its original answer and the character of its evidence, so the rule could be applied without re-researching anyone.
Assign the classification

Both tests pass: Practitioner-Researcher. Hands-on only: Practitioner. Study only: Researcher. Neither: Secondhand Explainer.
Rate confidence

High, Medium, or Low, with a one-line reason in the private notes. Confidence is recorded for every record and is not published per record.

Continuity is recorded, not scored

Continuity is not part of the classification. For every speaker, the earliest date of publicly verifiable AI-related work is recorded, along with whether it precedes Nov. 30, 2022, and it is reported alongside the classification rather than inside it.

The wording rule

Every output says no publicly verifiable evidence of, never no experience in. Absence of public evidence is not absence of experience, and the difference is the whole study. People who worked under confidentiality or deep inside a large organization can have years of hands-on AI work that leaves nothing a stranger can find.

Step three

The verification protocol

Ten to fifteen minutes per speaker. Every positive finding was recorded with its evidence URL and that page archived, so each coding decision can be defended after the fact.

What was checked, and where.
Check	What was looked for	Where
Listing and bio	Whether AI is the primary topic, expertise superlatives, stated AI claims, any fee band, any vendor disclosure.	The bureau listing and the speaker’s own site.
Hands-on work	Named products, deployments, AI roles, patents, code, dated training programs.	Speaker site, company sites, press, patent search, code hosts, and professional profiles viewed by hand.
Disciplined study	A research appointment, peer-reviewed AI papers, analyst reports, sustained AI reporting.	Scholarly search by hand, preprint servers, university pages, publisher archives.
Continuity	The earliest dated AI-related work, and whether the speaker’s own site mentioned AI before 2023.	Web archive snapshots of the speaker’s domain, dated publications, past event agendas.
Buyer evidence	Stated scope limits, a client or reference list, a described update process.	Speaker site, bureau listing, media kit.
Vendor relationships	Disclosed advisory, affiliate, investment, or employment ties to AI vendors.	Speaker site, bureau listing, professional profiles, press.

Access rules

Only publicly available information was used. No login-protected content was accessed.
No speaker and no bureau was contacted during collection. The study measures what a buyer can find on their own.
Scripted collection was limited to bureau category pages and web archive lookups, respecting robots.txt and site terms. Professional profiles and scholarly search were reviewed by hand and never scraped.
Pre-generative machine learning, data science, predictive analytics, and robotics work all count as AI-related work.
Every positive finding was recorded with its evidence URL and that page was archived, so each coding decision can be defended after the fact.
When a research session ran out of search budget mid-batch, the coder stopped and wrote nothing rather than falling back on a listing-only verdict. Thirty-four records produced that way on Sept. 17 were quarantined and re-researched, and several of those verdicts changed.

Superlatives, and what counted as one

Expertise superlatives were counted in each listing and bio, but only when used as unqualified self-description of AI expertise. A third-party award or ranking cited with a source does not count. The full list is published so a reader can see it rather than take the count on trust:

leadingtopforemostworld-renownedworld-classpremierpioneervisionary#1number onethe go-tomost sought-afterAI expertAI authorityAI guru

Step four

Quality control

Pilot

Two coders independently coded the same 20 profiles in week one, compared results, and refined the rubric wording before full collection began.
Unclear resolution pass

Records carrying an Unclear verdict were resolved in a single pass at the end of the run, against the written decision rules rather than case by case, so no verdict was left to stand as Unclear.
Author review

Every record was reviewed by the author against the published rubric after coding. Where a judgment was close, the governing rule was written down or tightened and every affected record was recoded, rather than the single case being decided on its own. That is why the decision log runs to twenty entries.
Data freeze

After adjudication the dataset was locked and the freeze date recorded. Any later change requires a logged reason and appears in the change log on the report.
Fact-check

Every published number was traced back to the frozen dataset by someone other than the analyst who produced it.

Agreement between coders

Coders: 1, with every record reviewed by the author

Agreement is published whatever it comes to. If it lands below the target, that is reported here and the affected findings are reported with that caveat attached rather than quietly dropped.

Step five

Data dictionary

Every field collected, and whether it can ever leave the research team. Fields marked private never appear in any output, in aggregate or otherwise. No record-level data is published, and no subgroup smaller than 10 speakers is reported.

Fields, values, and visibility.
Field	Values	Visibility
`record_id`	Sequential ID	Private
`capture_date, archive_urls`	Date; list of URLs	Private
`bureau_count`	How many included bureaus list the person	Aggregate only
`ai_primary_topic`	Yes / No	Aggregate only
`ai_topic_on_listing`	Yes / No, from the listing alone	Aggregate only
`superlative_count`	Count of unqualified expertise superlatives in the listing and bio	Aggregate only
`fee_band_published, fee_band`	Yes / No; the band as published	Aggregate only
`hands_on_test`	Pass / Fail / Unclear, with evidence type	Aggregate only
`study_test`	Pass / Fail / Unclear, with evidence type	Aggregate only
`peer_reviewed_ai_pubs`	0 / 1 to 2 / 3 to 9 / 10 or more	Aggregate only
`earliest_ai_evidence`	Year and month; source type	Aggregate only
`pre_chatgpt_evidence`	Yes / No / Unclear	Aggregate only
`site_ai_before_2023`	Yes / No / No archive available	Aggregate only
`full_length_video`	Retired Sept. 18, 2026. Collected, not reported, and not carried into the next edition.	Retired
`scope_limits_stated`	Yes / No	Aggregate only
`references_public`	Yes / No	Aggregate only
`update_process_described`	Yes / No	Aggregate only
`vendor_ties`	Disclosed / None found	Aggregate only
`source_type`	Practitioner / Researcher / Practitioner-Researcher / Secondhand Explainer / Insufficient evidence	Aggregate only
`confidence, coder_id, notes`	High / Medium / Low; initials; free text	Private

Guardrails

Ethics, privacy, and what gets published

Aggregate reporting only. The record-level dataset is private, access is limited to the research team, and it is retained for 24 months to support corrections and the next edition.
Only publicly available information was used. No speaker or bureau was contacted and no login-protected content was accessed.
No names, photographs, quotations, or paraphrases that could identify an individual or a bureau appear in any output.
Every percentage is published with its base, and no subgroup smaller than 10 speakers is reported at all.
The report had an independent fact-check, by someone other than the analyst who produced the numbers, and a legal review before publication.
Corrections are welcome from anyone, including speakers and bureaus whose listings were captured, and every correction that changes a published number appears in the dated change log on the report.

Known limitations

Public evidence understates real experience. People who worked under confidentiality or inside large organizations can have deep hands-on AI work that leaves no public trace. Every finding here is about what a buyer can verify, not about what a speaker has done.
Bureau AI category pages reflect bureau marketing choices, not a definitive population of AI speakers. Speakers who represent themselves are not listed by a bureau and so are not in this study at all, which is the single largest gap in its coverage.
Scholarly databases were unavailable during coding. OpenAlex, Semantic Scholar, Google Scholar and arXiv were blocked or rate-limited throughout, so peer-reviewed publication counts were assembled from CVs, anthologies, university pages and publisher listings instead. That almost certainly undercounts publications for some people, which would understate the study test rather than overstate it.
Coding is judgment applied to a rubric. One researcher coded every record and the author reviewed every record; where a call was close, the rule was written down and the affected records were recoded. A reader who disagrees with a judgment can see the rule it came from and the aggregate it lands in, but cannot re-open an individual record, because record-level data is never published.
Web archive coverage is uneven. "No archive available" is reported separately and is never treated as a negative finding.
The rubric measures the source of knowledge. It does not measure speaking quality, audience outcomes, or value for a given event.
The study is a snapshot of its capture window and will change with each edition.

Dropped before analysis: Public availability of a full-length recorded talk

A fourth buyer-evidence item was planned and tested: a publicly viewable, continuous talk of at least 20 minutes. It was dropped on Sept. 18, 2026, before analysis. Speakers commonly keep full talks private or host them somewhere that is not publicly searchable and send links to buyers directly, so what the measure captured was where a speaker hosts video rather than what a buyer can see. No figure is reported for it, because a low number would read as an absence of footage when it is really a hosting choice.

Repeatability

Running it again

The audit repeats every six months with the same rubric, reporting change over time from the second edition onward. Each edition keeps its own URL and its own numbers, so a citation of one edition is never quietly rewritten by the next.

What is published here is what another researcher would need: the population definition, the inclusion rules, the stratification, the seed, the two tests and their thresholds, the superlative list, the per-speaker time budget, the suppression rule, and the agreement target. The one thing not published is the record-level data, because it names people.

Edition: Fall 2026 edition
Brief prepared: September 17, 2026
Listings captured: October 1, 2026
Coding window: Sep 28, 2026 to Oct 2, 2026
Data freeze: October 2, 2026
Next edition: Spring 2027

Back to the findings

Author disclosure

Ryan Vet is a working keynote speaker, including on AI-related topics, and he conducted this study. He is not in it: he appears on none of the AI category pages captured, so there was nobody to exclude, and nobody was excluded. He is measured by the same rubric anyway, and publishes his own result against these tests on his site. This study is not a ranking, a directory, or an endorsement. No speaker and no bureau is named in it, and no bureau-level result is published.

Not a ranking, list, directory, or endorsement. No individual speaker is named in any output.
Not a judgment of any speaker’s quality, stage skill, or value. It measures publicly verifiable evidence only.
Not a bureau comparison. No bureau-level results are published.

Corrections

Corrections to the method or the findings are welcome, including from speakers and bureaus whose listings were captured. Every correction that changes a published number appears in the change log below with its date and reason.

research@ryanvet.com

Population and sample

One record, coded end to end

Inclusion rules

The classification rubric

The decision tree

Is there enough public information?

The hands-on test

The study test

Assign the classification

Rate confidence