Event Platforms

• Detection Priority — order recon.mjs probes
• Next.js / __NEXT_DATA__ — Stripe Sessions class
• Sessionize — public JSON API
• Lu.ma — JSON-LD Event block
• Eventbrite — JSON-LD Event block
• Custom / Markdown Fallback — last resort
• Adding a New Platform — contributor guide

Detection Priority

recon.mjs probes the event URL once via browse open + browse eval, then chooses the first matching platform from this list:

1. Next.js / __NEXT_DATA__ — document.getElementById('__NEXT_DATA__') returns a <script> tag.
2. Sessionize — <meta name="generator"> content matches /sessionize/i.
3. Lu.ma — location.hostname matches /lu\.ma/.
4. Eventbrite — <meta property="og:site_name"> matches /eventbrite/i.
5. JSON-LD Event — at least one <script type="application/ld+json"> block has @type === 'Event'.
6. Custom / markdown fallback — none of the above; the extractor uses browse get markdown and parses speaker blocks heuristically.

The order matters: Next.js sites often also embed JSON-LD, so probing __NEXT_DATA__ first gives us the structured data path before falling back to JSON-LD heuristics.

Next.js / __NEXT_DATA__

Detection signature

This script tag is emitted by every getServerSideProps / getStaticProps Next.js page. Its textContent is a JSON blob containing every prop hydrated into the React tree at build/request time — including the speaker list for an event microsite.

Extraction strategy — next-data-eval

recon.mjs walks the parsed JSON looking for arrays whose elements are objects with both a name-ish key AND a linkedin substring somewhere. It records the JSON path of every such array (e.g. .props.pageProps.featuredSpeakers.speakers.items) into recon.nextDataPaths. The Phase B extractor then runs ONE browse eval to harvest those arrays and union them into a single people list.

Sample output shape (Stripe Sessions 2026)

A typical speaker object inside one of those arrays:

Known gotchas

• The walker also matches talks[N].speakers arrays (a denormalized re-listing of the same speakers per session). recon.mjs filters those out via regex so we don’t double-count.
• Some Next sites lazy-load speakers AFTER hydration. The 2.5s browse wait timeout is usually enough; if a site fails extraction, bump the wait to 5s before declaring it a different platform.
• Field names vary across sites (companyName vs company vs org, linkedInProfile vs linkedinUrl). The Phase B extractor normalizes via fallback chains.

Sessionize

Detection signature

Extraction strategy — sessionize-api (stub in v0.1; full implementation in a future phase)

Sessionize exposes a public read-only JSON API at https://sessionize.com/api/v2/{event_id}/view/Speakers. The event ID is in the page URL or embedded JS. v0.1 of recon.mjs only sets strategy: "sessionize-api" and emits the URL — no API discovery yet.

Sample output shape

Known gotchas

• Some Sessionize-hosted pages bury the event ID behind a custom domain. We may need to scan the page for sessionize.com/api/v2/... URLs and pull the ID from there.
• The Sessionize API returns a flat speaker list — much cleaner than scraping Next.js, when it works.

Lu.ma

Detection signature

Lu.ma always serves on lu.ma (or rarely a custom CNAME); the hostname check is decisive.

Extraction strategy — json-ld (stub in v0.1)

Lu.ma embeds an Event JSON-LD block with attendee/speaker info, but the volume of structured data is event-dependent. v0.1 sets strategy: "json-ld" and defers actual extraction to Phase B+. The fallback markdown extractor handles Lu.ma pages reasonably well in the meantime.

Sample output shape

Known gotchas

• Lu.ma often gates the full attendee list behind a login. Public-facing pages usually only show “featured” speakers, not the full list.
• Some Lu.ma events are private; recon.mjs should NOT crash if __NEXT_DATA__ isn’t present and JSON-LD is empty — it falls through to the markdown strategy.

Eventbrite

Detection signature

Extraction strategy — json-ld (stub in v0.1)

Eventbrite emits standard Event JSON-LD on every public event page. Speakers are usually only in the prose body, not the structured data — Eventbrite is more of an RSVP platform than a speaker-directory platform. We may need to combine JSON-LD parsing with markdown extraction of the description body.

Sample output shape

Known gotchas

• Eventbrite event pages can be heavy. If browse cloud fetch fails or returns thin content, use browse get markdown instead.
• Most Eventbrite events do NOT publish a speaker list at all. This platform is low-yield for prospecting.

Custom / Markdown Fallback

Detection signature

None of the above match. recon.mjs sets platform: "custom" and strategy: "markdown".

Extraction strategy — markdown

The Phase B extractor calls browse get markdown, splits the output on heading boundaries (####, ###, ##), and treats each block as a candidate speaker:

• Line 1: name (must start with capital letter)
• Line 2: title/role
• Line 3: company
• Anywhere in the block: linkedin.com/in/{handle}

This is a best-effort fallback. Coverage is typically 60-80% of the actual speaker list; field accuracy is lower than the structured paths.

Sample output shape

Known gotchas

• Many static event sites use cards (not headings) for speakers. The heading-split heuristic misses those entirely.
• “Section title” lines (## Speakers, ## Schedule) get parsed as candidate speakers and need to be filtered downstream by checking for the LinkedIn pattern.
• If markdown extraction yields zero people, the pipeline should surface a “platform unsupported” error rather than emit an empty people.jsonl.

Adding a New Platform

1. Add a detection branch in recon.mjs probe() after the existing ones. Pick a cheap signal (a meta tag, hostname, or a specific script tag) that’s distinctive.
2. Pick a strategy name — kebab-case verb-phrase like sessionize-api or json-ld-events.
3. Add an extractor branch in extract_event.mjs that handles the new strategy.
4. Add a section to this file with: detection signature, extraction strategy, sample output shape, known gotchas.
5. Drop a fixture under scripts/__fixtures__/{platform}-snapshot.json capturing the expected recon.json shape so future refactors don’t silently break extraction.

Keep the detection branches small. If a platform needs more than ~20 lines of detection logic, factor it out into scripts/detectors/{platform}.mjs and import.