# The Speaker-Program Teardown: Five Agendas to a Ranked, Current Slate

*Turn any event's current and archived agendas into a ranked slate where every row has the talk, its slot type, two dated employer sources, and a first line taken from the talk.*

- Canonical URL: https://www.refolk.ai/guides/speaker-program-teardown-ranked-slate
- Pillar: Recruiting and sourcing
- Format: Teardown
- Published: 2026-11-09
- Last reviewed: 2026-11-09
- Reading time: 21 min
- Keywords: how to source candidates from conference speaker lists, conference speakers as recruiting leads, find candidates from event agendas, sourcing passive candidates from conference talks, recruiting outreach referencing a conference talk

## Key takeaways

- At KubeCon Europe 2021, 95 accepted talks had 133 speakers, about 1.40 per talk, so counting talks instead of people undercounts speakers by roughly 40%.
- KubeCon's schedule labels paid keynotes "Keynote Sessions Sponsored". Rank those speakers below CFP speakers, who competed in a pool that accepts roughly 15% of submissions, but keep them on the list.
- If you assume people leave at a steady rate and take the 4.1-year median tenure, about 15.5% of speakers will have changed employer one year after their talk and about 39.8% three years after, so confirm every current employer with two dated sources.
- Refolk's index has 425 people with Platform Engineer titles in Germany and 34 with Developer Advocate or Developer Relations titles, a 12.5x gap. On a German agenda, cutting vendor evangelists removes only a small, concentrated group.
- The UK regulator (the ICO) says that when data comes from a public source, the privacy notice is due at first contact at the latest and never more than one month after collection, so log the collection date on every slate row.
- No published study measures the reply lift from mentioning someone's conference talk, so test it against a control group yourself rather than quoting the 35 to 50% figures that circulate without a sample.

This guide shows how to turn a conference's speaker programs, the live one and the archived ones, into a slate you can message this week. The finished slate has one row per real person, split by who chose the talk, checked against where the person works now, and ranked on what they said on stage. It is for in-house recruiters, sourcers, talent leaders and founders with a live specialist search. The worked example is a platform engineering search in Germany, sourced from five editions of KubeCon. I carry it through every cut with the published counts, the ratios that predict your own counts, and the wrong turns that cost the most time.

## What does a speaker-program teardown produce?

A speaker-program teardown produces a ranked slate in which every row carries the talk title, the talk date, the slot type, current-employer evidence from two dated sources, and a first line built from the talk itself. Anything less is a list of names next to logos that may be out of date.

Most public advice on how to source candidates from conference speaker lists stops at "speaker lists are gold", or at a one-off scrape of this year's program. That skips the parts that decide whether the list is usable: past editions, slots that sponsors paid for, co-speakers and panelists, and the people who have changed employer since they spoke.

### The worked example

- **The brief.** A platform engineer in Germany who has built and run internal Kubernetes platforms, not someone whose job is selling one.
- **The event.** KubeCon + CloudNativeCon Europe. Its tracks match the must-haves, and its organiser, CNCF, publishes more about how talks get selected than most events do.
- **The window.** Five editions: the current program and four archived ones.
- **The market.** In Refolk's index of professional profiles, 425 people in Germany hold Platform Engineer titles. In a 25-profile sample, 5 work at IBM. Speakers are a small, self-selected slice of that pool, which is why they are worth the effort.

### Terms used throughout

- **CFP slot:** a talk chosen through the call for papers, an open submission process judged by a program committee.
- **Sponsored slot:** a speaking slot that came as part of a sponsorship package.
- **Maintainer session:** a session in which the maintainers of an open-source project talk about that project.
- **Vendor evangelist:** a developer advocate or developer relations (DevRel) person speaking about their own employer's product.
- **Edition:** one year's instance of the event, with its own program page.

I am not publishing the per-person counts from my own run of this example. The published event figures below are real. The ratios built from them tell you what to expect at each cut. The count log in the procedure section is where your own numbers go.

## How many speaker rows survive each cut?

Raw speaker rows shrink at four predictable points: resolving and merging people, moving sponsored slots down, flagging vendor evangelists, and checking who has changed employer. CNCF publishes enough about KubeCon to estimate the first ratios before you scrape a single page.

| Edition | CFP submissions | Accepted | Speakers |
|---|---|---|---|
| EU 2021 | 621 | 95 talks | 133 |
| NA 2020 | 856 | n/a | 345 |
| NA 2023 | 1,871 | 359 sessions (88 maintainer) | 554 |

The EU 2021 figures come from CNCF's write-up of its schedule selection, and the NA 2023 figures come from CNCF's transparency report. Acceptance was 15.3% for EU 2021 and 15% for NA 2020. CNCF describes its average acceptance rate as being in the low to mid-teens.

### What these numbers predict for your run

- **Co-speakers.** EU 2021 accepted 95 talks with 133 speakers, or 1.40 speakers per talk. If you count talks, you undercount people by about 40%. NA 2023 works out to about 1.5 speakers per session. Many agendas print co-speakers as a single string such as "Name and Name, Company", so the first job is to split those strings into separate people.
- **Repeat speakers.** Sessionize, the call-for-papers platform, reports that 27% of its speakers applied to two or more events. That measures applications across different events, not repeat talks at one event. Treat it as a sign that repeats are common, not as a rate to plug into your log.
- **Maintainers.** 88 of NA 2023's 359 sessions were maintainer sessions, which is 24.5%. Maintainer sessions are strong evidence of hands-on work, but many maintainers work for vendors, so they still go through the next cut.
- **Scale.** The Barcelona edition of KubeCon Europe advertises 350+ sessions and 700+ speakers. Five editions at that size give you raw rows in the thousands before any dedupe.

#### KubeCon Europe 2021, from submission to end-user talk

| Stage | Figure | Note |
| --- | --- | --- |
| CFP submissions | 621 | from 910 proposed speakers at 304 companies |
| Accepted talks | 95 | 15.3% acceptance, 133 speakers |
| End-user talks | 17 | 17.89% of accepted; vendors gave about 62 |

*Of 621 CFP submissions at KubeCon Europe 2021, 95 were accepted and only about 17 of those came from end-user companies.*

**1.40 - Speakers per accepted talk, KubeCon Europe 2021**

133 speakers across 95 talks. A count of talks understates the number of people on the program by about 40%.

A sanity check for your log: if your unique-human count comes out at or below your talk count on an event with co-speakers, you probably merged co-speaker strings into single rows instead of splitting them.

## Which slots were peer-selected and which were bought?

Read the slot label on the archived schedule, not the size of the name on the page. On KubeCon's live schedule, a paid keynote carries the label "Keynote Sessions Sponsored" and an unpaid one carries plain "Keynote Sessions".

Sponsored does not mean unvetted. CNCF's sponsor terms say: "Session content must meet CFP requirements and is subject to approval by CNCF. No sales or marketing pitches." A later sponsor prospectus repeats that sponsor speaking slots must meet CFP criteria and need program committee approval. It also lists an attendee contact list as part of the package, which tells you what the sponsor was buying.

So a sponsored talk passed a quality bar but did not compete in the pool where roughly 15% of submissions get through. My rule is to rank sponsored speakers below CFP speakers and keep them on the slate.

The split is not always published. The NA 2023 transparency report lists 20 keynotes "includes sponsored keynotes" and gives no sponsored count. Count the labels yourself, edition by edition, and log the number you moved down a tier.

### Vendor employees versus vendor evangelists

In EU 2021, 65.26% of accepted talks came from vendors and 17.89% from end users, the companies that run the software rather than sell it. Cutting every vendor employee would throw away most of the program, including engineers who built the thing they are presenting. The cut is narrower:

- **What the flag is:** the speaker holds a DevRel title at the vendor whose product is the subject of the talk.
- **What it proves:** the talk is part of the speaker's job of promoting a product, so it says little about whether they have run that product in production for someone else.
- **When it lies:** some people move into DevRel after years running production platforms, and their talks show it. A DevRel title therefore triggers a check of the recording, not an automatic removal.

| Market | Title (headline match) | Profiles |
|---|---|---|
| US | Developer Advocate / Relations | 298 |
| Germany | Developer Advocate / Relations | 34 |
| Germany | Platform Engineer | 425 |

All three counts are from Refolk's index, using a match on the profile headline title.

**12.5x - Platform Engineers per DevRel title in Germany, Refolk's index**

425 Platform Engineer titles against 34 Developer Advocate or Developer Relations titles.

For the worked example, German DevRel is a thin and concentrated pool. In the 25-profile sample, JetBrains has 3 and SAP has 2. On a German agenda, the evangelist flag removes a small group of people you can recognise by name. The US pool is 8.8x larger, so on US-heavy agendas expect the flag to remove a larger share. Remember that these are counts of profiles, not of speakers. Use them for proportions, not for predicting your exact numbers.

#### Where each speaker lands on the slate

Horizontal axis runs from Sponsored or evangelist slot to Peer-selected CFP talk. Vertical axis runs from Employer unverified to Two dated current sources.

| Quadrant | What it means |
| --- | --- |
| Verified, sponsored | Keep in a lower tier; check the recording for real hands-on work |
| Verified, CFP | Top of the slate; message this week |
| Unverified, sponsored | Park unless the pool is thin |
| Unverified, CFP | Verify first; if they moved, follow them to the new employer |

*Who chose the talk sets the ceiling on a speaker's rank, and how well the current employer is evidenced decides whether you can message them this week.*

## How stale is the employer on the speaker badge?

The company printed on a speaker page was true on the day of the talk and nothing more. On a tenure-based estimate, about 15.5% of speakers will have changed employer a year later and about 40% three years later.

The U.S. Bureau of Labor Statistics' most recent tenure release puts median tenure with the current employer at 4.1 years for all wage and salary workers. In its previous release, the median for workers aged 25 to 34 was 2.7 years. Assuming people leave at a constant rate, the share who have moved after t years is 1 - 0.5^(t / median):

| Years since talk | Median 4.1y (all workers) | Median 2.7y (ages 25 to 34) |
|---|---|---|
| 1 | 15.5% | 22.6% |
| 2 | 28.7% | 40.2% |
| 3 | 39.8% | 53.7% |

These are rough expectations, not facts about your speakers. Median tenure in the current job does not strictly support a constant departure rate. Other BLS figures push in both directions. Median tenure in management, professional and related occupations is longer, at 4.9 years, which lowers the estimate. But 22% of wage and salary workers had been with their current employer a year or less, and younger speakers move faster. Treat the table as a reason to check every row, and measure your own moved fraction rather than quoting these.

BLS updates these figures every two years. To re-check them, look at the "Employee Tenure" release and rerun the formula with the new median.

> **Rule:** Two dated sources, or the row is unknown
>
> No current employer goes on the slate without at least two dated sources that agree. One source, or two undated ones, makes the row "unknown", and unknown rows are never messaged at the badge company.

### The archive decays on the same curve

Old program pages disappear at the same time as the people move on. Pew found 8% of pages from 2023 and 15% from 2022 were no longer accessible, rising to 38% for pages from 2013. Across all pages that existed from 2013 to 2023, 25% were gone. The Internet Archive estimates the Wayback Machine has rescued roughly 15% of those otherwise dead pages.

Nobody has published survival rates for speaker pages specifically, or for how often a company field is lost when a page is archived. Measure both in your own run. The result for planning is that editions three to five years old cost the most verification per usable row: more pages are dead, and more of the people on them have moved.

The employer check is where the hours go: an estimated five minutes per person, multiplied across several hundred people. In [Refolk](/) you can ask in plain English for the speakers at a given edition whose employer has changed since their talk. Each current employer comes back attached to the record it came from, so every row starts with one dated source and you only need to find the second.

## The procedure, edition by edition

The procedure runs in eight steps, from the brief to a sent message, and logs a count at every cut so each loss can be explained. Time estimates are rough and unverified. Budget your own after the first edition.

#### From agendas to a ranked, messaged slate

1. **Write the brief and choose the event** - Sit with the hiring manager, write down the must-haves, and pick one event whose tracks match them. Done when both are written down.
2. **Enumerate every edition** - Check the live site, the /archive/YYYY/ paths and the Wayback CDX server. Done when you have one URL per edition, flagged live, archived, recovered or dead.
3. **Extract one row per speaker per session** - Record talk title, talk date, slot label and listed company for every speaker in every session. Done when the raw row count is logged.
4. **Resolve and dedupe the humans** - Split co-speaker and panel strings, then merge the same person across sessions and editions using a Sessionize URL or name plus company plus title. Done when the unique-human count is logged.
5. **Cut sponsored slots and flag vendor evangelists** - Move Sponsored slots to a lower tier, then flag DevRel speakers at the vendor whose product the talk covered. Done when both counts are logged.
6. **Check the current employer with two dated sources** - Find at least two dated sources for each person's current employer, at an estimated five minutes per person. Done when moved, stayed and unknown are counted.
7. **Score brief fit from the recording and rank** - Score what each talk actually covered against the brief, ignoring the bio, and agree the order with the hiring manager. Done when the ranked slate is signed off.
8. **Send the first line with the privacy notice** - Write each opener from a specific point in the talk and include the Article 14 notice in the same message. Done when each row is sent within one month of its collection date.

### Notes on the steps that go wrong most

- **Step 2, finding editions.** The Linux Foundation keeps past KubeCon editions under paths like `events.linuxfoundation.org/archive/2024/`. For anything missing, query `web.archive.org/cdx/search/cdx` with the program URL, `matchType=prefix` and `filter=statuscode:200`. The CDX server returns 7 fields by default and caps a single query at 150,000 results, so query each edition separately rather than the whole domain.
- **Step 3, structured sources.** Many events run their call for papers on Sessionize, which reports 321,000 speakers applying through it, 88,000 public speaker profiles and 12,300 events. Public profiles expose the name, tagline, company, bio, topics, sessions, past and upcoming events, and social links. That makes the profile URL the best dedupe key you will get.
- **Step 7, scoring.** CNCF publishes KubeCon session recordings on its YouTube channel after each event. Score 0 to 3: 3 means the speaker describes running the thing in production with specifics that match the must-haves; 2 means hands-on but with a different stack or scale; 1 means an overview; 0 means a product pitch or off-brief. Note the timestamp that justified the score, because the first line comes from it.

The two published methods disagree on order. One pulls each speaker's work history before scoring and treats the bio as "marketing copy". Another cross-references speakers against target profiles before the event happens. For a live search on past editions, I check the employer before ranking, because the ranking is wasted on someone who is no longer reachable at the company that matters.

**Teardown count log and slate row**

```
COUNT LOG
Event:
Editions (URL :: live / archive / recovered / dead):
Raw speaker rows (one per speaker per session):
Unique humans after split and merge:
Sponsored slots moved to lower tier:
Vendor evangelists flagged:
Employer check (stayed / moved / unknown):
Brief fit 2 or 3:
Collected on (date):

SLATE ROW
Rank:
Name:
Dedupe key (Sessionize URL, or name + company + title):
Talk title:
Talk date and edition:
Slot type (CFP / Maintainer / Keynote / Keynote Sponsored / Sponsored):
Company listed at talk:
Current employer:
Evidence 1 (source, date seen):
Evidence 2 (source, date seen):
Brief fit (0 to 3) and recording timestamp:
First line:
Collected on:
Notice sent on:
```

*Keep one log per search and one slate row per human. Fill in every line; a blank count is a cut you cannot explain later.*

I ran this search: `Platform engineers in Germany who gave a CFP talk at KubeCon Europe between 2022 and 2026, excluding sponsored sessions, with current employer` - [see the full result list](https://www.refolk.ai/s/pnvsabjs6c).

*You get speakers on the CFP track with their talk and current employer attached, ready for the second dated source and the recording check.*

## How speaker-list sourcing goes wrong

Speaker-list sourcing fails in eight known ways, and most of them produce false positives: people who look like strong leads but are the wrong person, the wrong kind of speaker, or no longer at the company. Three of them are the wrong turns this example is built to catch.

> **Watch out:** The badge is a date, not an address
>
> A message sent to someone at the company printed on a talk page from two years ago has a tenure-based chance of nearly 30% of reaching a person who has left. Do not send until two dated sources agree on the current employer.

### The three wrong turns in the worked example

1. **Ranking on keynote fame.** The tempting first move is to put the keynote speakers at the top. A keynote may be a "Keynote Sessions Sponsored" slot: a well-known executive whose vendor paid for the stage. Check the slot label on the archived schedule before any keynote gets a rank.
2. **Trusting the bio over the talk title.** Bios are written by the speaker for promotion, and one published method calls them "marketing copy". The false positive is an inflated title or a scope the talk does not support. Check the claim against the talk title, the recording and a dated work history.
3. **Messaging the badge employer.** Even a one-year-old talk carries an expected move rate of about 15.5%. Check for two dated current sources and follow anyone who has moved to their new company.

> Rank on what they said on stage, verify where they sit now, and ignore how big the name was printed.

### The other five

- **Name-only dedupe.** Common names merge different people, and panelists get duplicated across rows. Match name, company and title together, or use the Sessionize URL. If unsure, leave the records separate.
- **Vendor evangelists counted as practitioners.** 65% of EU 2021's accepted talks came from vendors. Ask whether the speaker's employer sells the product being presented, and check whether the speaker's title is DevRel.
- **Archive gaps read as absence.** At 38% decay for ten-year-old pages, a missing edition does not prove someone did not speak. Query Wayback CDX before you drop a row or conclude a person is a one-time speaker.
- **Late notice.** The one-month GDPR clock starts at collection, not when you send. Log the scrape date per row, not per search.
- **Unbacked personalization claims.** The "35 to 50%" reply lifts quoted for personalized outreach cite no sample. Do not use them as benchmarks or promise them to a hiring manager.

#### How much each part of a speaker record proves

1. **Bio** - Self-written promotional copy; proves how the speaker wants to be seen
2. **Company on the badge** - True on the talk date only; decays roughly with tenure
3. **Slot label** - Proves who chose the talk: peers, a sponsor, or the project
4. **Talk title and abstract** - Proves what the speaker claimed the talk would cover
5. **Recording** - Proves what they actually said, with a timestamp you can quote

*The bio is the outer, weakest layer of a speaker record and the recording is the inner, strongest one, so rank from the inside out.*

## Writing the first line from what they said on stage

Build the first line from a specific point in the recording, quoted or closely paraphrased with its timestamp. Treat any reply lift from it as unproven until you measure it, because no published study isolates the effect of mentioning a talk.

The nearest evidence measures something else:

- **Individual versus bulk.** LinkedIn reports that individually sent InMails see response rates roughly 15% higher than bulk sends. It counts any response within 30 days and does not disclose a sample size.
- **Hand-typed versus AI-drafted.** In a study of more than 5 million messages, recruiter-typed first emails got replies at 12.6%, about 2.5x the rate for AI-drafted emails.

Both results reward effort spent on one person. Neither shows that mentioning a talk is what earns the reply. That is still the honest bet for recruiting outreach that references a conference talk: a line only that speaker could receive is individual by construction. But run it against a control group of generic first lines and record the result in your count log.

Avoid praising a talk you have not watched. "Loved your talk" with no detail reads as bulk sending.

### The privacy notice goes in the same message

The ICO applies the Article 14 notice deadline to data taken from a publicly accessible source. The latest point to give notice is when you first communicate with the person, and the one-month limit from collection still applies. The ICO also warns that "you should not assume that data protection law doesn't apply because the data is publicly available". The EDPB, which coordinates EU data protection regulators, has not published guidance specific to speaker pages, so take advice locally if you source EU-based speakers.

> **Note:** Check your lawful basis with counsel
>
> A secondary source reports that recruitment is not on the UK's list of "recognised legitimate interests" in the Data (Use and Access) Act 2025, so a recruitment search still needs a legitimate interests assessment. I have not confirmed this against the primary text; confirm it before you rely on it.

**First message with a talk-based opener and privacy notice**

```
Subject: Your KubeCon talk on {topic}

Hi {first name},

I watched "{talk title}" from {event, month and year}. Your point at {timestamp} about {specific claim, in their words where possible} is the problem the platform team at {company} is working on now.

I am hiring a {role} in {location}. {One sentence on scope that ties back to the claim above.} Would you be open to a short call?

How I found you: your name and talk come from the public {event} program, and I confirmed your current role on {source}. I hold your name, talk details, current employer and the date I collected them, for this search only. You can ask what I hold, or ask me to delete it, at {contact}. Full notice: {privacy notice link}.

{Your name}
```

*Write line one from the recording, not the abstract, and keep the notice paragraph intact.*

LinkedIn Recruiter is reported to restrict InMail for accounts whose response rate falls below 13% on 100 or more sends within 14 days. That figure comes from a secondary source and I have not confirmed it against LinkedIn's own documentation. Either way, it argues for sending this slate in small batches of well-checked rows rather than all at once.

## Keeping the slate current after the first send

A speaker slate goes out of date on two clocks: the event publishes a new edition each year, and the people on it keep changing jobs. Re-run the teardown when each new program goes live, and re-check employers on any row you have not contacted within a quarter.

- **New editions.** Add the new edition as one more row in the count log rather than starting over. The dedupe key carries repeat speakers forward, and a person who has spoken more than once is worth noting on their row.
- **Your own rates.** After two or three searches, replace the borrowed figures with your own: the share of rows lost to dedupe, the share of employers that moved per year since the talk, and the reply rate for talk-based openers against your control group. Those are the numbers that will predict your next run.
- **Movers.** Anyone flagged "moved" is a lead at their new company, and the talk is still a valid opener. When the slate needs topping up, Refolk can return the same speakers with their current employer, so a stale edition becomes fresh rows without a second round of manual searching.

#### Before you call the slate done

- [ ] Every edition in the window has a URL flagged live, archived, recovered or dead in the count log
- [ ] The log shows raw rows, unique humans, sponsored moves, evangelist flags, and moved, stayed and unknown counts
- [ ] No two people were merged on name alone
- [ ] Every keynote's slot label was read on the archived schedule, not assumed
- [ ] Every ranked row has two dated sources that agree on the current employer
- [ ] Every brief-fit score of 2 or 3 cites a recording timestamp, not the bio
- [ ] Every first line refers to a specific point in that person's talk
- [ ] Every row has a collection date, and its message, with the notice, goes out within one month of it
- [ ] A control group of generic first lines is set aside so the talk-based opener can be measured

## Frequently asked questions

### Are sponsored conference talks worth sourcing from?

Yes, but rank them below peer-selected talks. At KubeCon, sponsored sessions must meet the CFP's content requirements and need organiser approval, so they clear a quality bar. They did not compete in the call for papers, which accepts talks at a rate in the low to mid-teens. Read the slot label on the archived schedule and give sponsored speakers their own lower tier rather than deleting them.

### How many past editions of a conference should I pull for a search?

Three to five editions is a workable window, and each older one costs more to verify per usable row. Pew found 8% of one-year-old pages and 38% of ten-year-old pages had gone dead. A tenure-based estimate puts employer moves near 40% three years after a talk. Pull the older editions when the pool is thin, and plan extra time for the employer check.

### Do I need to send a GDPR notice when sourcing from a public speaker list?

Under UK GDPR, yes. The ICO applies Article 14 to data taken from a publicly accessible source. Privacy information is due when you first communicate with the person at the latest, and the one-month limit from collection still applies. The ICO also warns not to assume data protection law stops applying because data is public. Log the collection date on every row.

### Does mentioning someone's conference talk improve reply rates?

Nobody has published evidence either way. LinkedIn reports that individually sent InMails get roughly 15% more responses than bulk sends. A study of more than 5 million messages found recruiter-typed first emails got replies at 12.6%, about 2.5x the rate for AI-drafted ones. Neither study isolates a talk reference. Test it yourself against a control group.

### How do I dedupe conference speakers across several years of agendas?

Use a stable key, not the name. A Sessionize vanity URL is the best key where one exists. Otherwise, require name, company and title to match together before merging. Split co-speaker and panel strings into separate people first. If a match is uncertain, leave the records separate: two rows for one person costs less than one merged row covering two people.

### What if an old edition's speaker page no longer exists?

Query the Wayback CDX server before you treat the edition as lost. Use the program URL as a prefix and filter to status 200 captures. The Internet Archive estimates the Wayback Machine has saved about 15% of pages that would otherwise be dead. Mark the edition as recovered or dead in your log, and never read a missing page as proof that someone did not speak.

---

*From the Refolk guide library. I revise these guides rather than replacing them, so the current version is always at https://www.refolk.ai/guides/speaker-program-teardown-ranked-slate*