<- all tokdocs
2026.10.02@build_in_public

OpenAI's plugin guidelines ban the keyword-matched description this video teaches, and ChatGPT was already recommending plugins inside conversations by September 4, 25 days before the announcement the video credits

AI
Watch on TikTok

View on TikTok ->

This is a 1 minute 59 second clip at 1080x1920, posted 2026-10-02 at 00:58 UTC by @build_in_public (channel nickname "Build in Public"), carrying 81,200 views, 3,651 likes, 789 reposts, 43 comments and 3,967 saves at capture, scored to the creator's own "original sound" track, and tagged #searchengineoptimization #seo #leadgen #digitalmarketing. I read 28 of the 60 extracted frames and all 337 transcript words. The clip opens on a bearded man in a blue graphic tee sitting under a boom-arm microphone and a ring light, with burned-in white captions reading "This is what you should spend the next week playing around with." From frame 2 onward the shot becomes a handheld phone recording of a laptop screen, finger pointing at the text, with the browser address bar visible at the top reading "...m/gregisenberg/status/2105454208040206773?s=46" and a bookmarks bar showing Drive, Gmail, LinkedIn and X icons. The screen shows an X post page headed "Post" from "GREG ISENBERG" with a blue verification check and the handle "@gregisenberg". The post body, readable across frames 7 through 34, says: "I literally can't think of a better risk/reward bet right now than building a ChatGPT plugin and getting in front of their 1.2 billion weekly users. / As of yesterday, ChatGPT recommends plugins in the middle of conversations. Millions of questions get asked every day with no plugin to recommend. Each one is an empty slot. / Facebook apps in 2007. iPhone apps in 2008. TikTok in 2019. Every big platform has a few months where distribution is free, and most people miss it. / I saw this firsthand when I lived in SF building venture backed companies. They are sacred times! / If I were starting this week: / 1. Find a tool or workflow people already Google for or do by hand every week (PDF to Excel, invoice generator, weekly client reports) / 2. Build it as a ChatGPT plugin / 3. Write the description using the exact phrases people type into ChatGPT / 4. Ship 5 of them and see which ones get picked up / 5. Double down on the one with the most traction / In 12 months these slots will be crowded but right now they're pretty empty. / Most people will realize ChatGPT plugins were the biggest free distribution channel of 2026 sometime in 2027." Below that sits a quoted post from "Pietro Schirano" with a blue check and handle "@skirano · Sep 30" reading "Stop what you're doing and build a plugin extension for ChatGPT. We've grown over 2000% since the announcement!" Frame 39 shows that quoted post's attached chart: a bar series with x-axis ticks labeled 2026-09-21, 2026-09-24, 2026-09-27 and 2026-09-30, flat through September 27 and spiking across the last three bars, with the post footer reading "Sep 30, 2026 · 1M Views" and no y-axis label or units anywhere on the image. Frames 40 through 49 scroll into the reply thread: "Greg Thomas" with handle "@gregscribbles" asking "but how am I monetizing this? I don't g[et it]" stamped "5 PM · Sep 30, 2026 · 4,743 Views"; a reply from "Ezequiel Costales | AI CRO for Ecommerce" at "@ezzcosta · 2h" reading "Simple: it's top of funnel exposure. / Huge awareness generation opportunity to generate value and convert the ones who look you up / A ton if not most apps require you to have an account with them so that itself Is a way to bring people into your world" with 5 likes and 672 views; and a reply from "@PumpkinSpiced7 · 3h" ending "...unless you already have a produ[ct]" with 17 likes. Frame 51 returns to the talking head for "So this looks like a great lead gen opportunity," and frame 60 closes back on the Isenberg post with the caption "And shout out to Greg Isenberg for this post."

The in-conversation recommendation did not start on September 29, and developers were complaining it had broken three weeks earlier

Isenberg's post, visible verbatim in frames 7 through 13, says "As of yesterday, ChatGPT recommends plugins in the middle of conversations." The narrator compresses this to "As of two days ago." Neither framing survives a check against OpenAI's own developer forum. A thread titled Some plugins are not found in ChatGPT conversations since Sep 9, running from September 11 to October 4, 2026, has multiple plugin developers reporting that their plugins had been surfacing inside conversations and then stopped. One developer reports their plugin was "Suggested in conversations daily 09-04 to 09-09." Another says theirs was "recommended in chats for nine days before that." A third says discovery was "working normally until Sep 18." In-conversation plugin surfacing was running at least 25 days before DevDay and had already shipped, broken and been reported on by the time Isenberg described it as a one-day-old capability.

What OpenAI actually announced on September 29, 2026 is narrower. The Every DevDay recap renders it as "ChatGPT will soon be able to recommend relevant plugins during a conversation, and the app will include a directory where you can browse them," which is a future tense. TechCrunch's September 29 report describes "improved ranking and recommendation systems within its directory" plus "a redesigned flow that offers clearer feedback" for submissions. The DevDay 2026 announcement list uses OpenAI's own phrasing, "improved ranking and recommendations in the directory and in conversations." Improving ranking on an existing surface is not opening an empty one.

The 1.2 billion figure checks out as stated. Simon Willison's live blog of the keynote records Sam Altman saying "1.2B people use ChatGPT every week," and notes the opening slide showed 1.2 billion ChatGPT users alongside 2.5 million businesses. That number is ChatGPT's total weekly user base. It is not a count of people who have ever opened a plugin, and no source I found publishes that second number.

OpenAI's published plugin guidelines prohibit step 3 of the five-step playbook

Step 3 on screen in frames 22, 28 and 31 reads "Write the description using the exact phrases people type into ChatGPT." The narrator underlines it at 0:50: "This is just SEO, by the way. Know your bottom of funnel keywords."

OpenAI's plugin guidelines rule this out directly. Plugins must not include "descriptions, titles, tool annotations, or other model-readable fields" designed to manipulate how the model selects plugins or to "interfere with fair discovery." The same page says plugin names should avoid being "overly generic or single-word dictionary terms," which is exactly what a keyword-first naming strategy produces when the target query is "pdf to excel" or "invoice generator." It also requires that "Plugin names and descriptions must be clear, accurate, and straightforward," and bars descriptions that compare against competitors or make unverifiable claims.

The underlying mechanic is also different from search SEO. The plugins developer docs define a plugin as a package combining skills, MCP servers and an optional UI, with the model reading tool annotations to decide when to invoke it. Writing those model-readable fields to match consumer search phrasing is the specific behavior the guidelines name as manipulation. A creator following the video's step 3 literally is writing the thing a reviewer is told to reject.

Pietro Schirano's 2,000% came from an OpenAI launch partnership that predates the announcement by four weeks

The video treats Schirano as proof that the playbook works, saying at 1:03 "Pietro Schirano is doing this." The quoted post in frames 25 through 58 reads "Stop what you're doing and build a plugin extension for ChatGPT. We've grown over 2000% since the announcement!" The chronology runs the other way.

Schirano is founder and CEO of MagicPath, which raised a $6.6M seed round led by Khosla Ventures and Abstract in May 2025 and later took a $2M insider round at a $100M valuation. His own post history shows MagicPath was already live as a ChatGPT plugin across web and desktop, credited to "our friends at OpenAI," in a post whose snowflake ID resolves to September 2, 2026. On DevDay itself he posted MagicPath directly in ChatGPT's sidebar "in collaboration with OpenAI", timestamped September 29, 2026. The 2,000% post came the next day, September 30.

So the sequence is a funded design-tool company with a named OpenAI collaboration shipping a sidebar integration on announcement day, then posting a spike chart the day after. That is a launch-partner result. The video presents it as the outcome a person who finds an empty keyword slot and ships five plugins can expect. The chart in frame 39 also carries no y-axis label, no units and no baseline value, so 2,000% is a percentage of an unpublished starting number.

I could not fetch Schirano's posts directly because x.com returns HTTP 402 to automated requests. The post URLs above appear in search results with their text, and the dates come from decoding the X snowflake IDs, which is deterministic. Treat the exact post timestamps as derived rather than read off the page.

The monetization answer quoted in the clip describes a funnel OpenAI's rules currently block

At 1:08 the narrator reads the reply from @ezzcosta, visible in frames 43 through 46: "it's top of funnel exposure. Huge awareness generation opportunity to generate value and convert the ones who look you up. A ton if not most apps require you to have an account with them so that itself Is a way to bring people into your world." The narrator concludes at 1:42, "So this looks like a great lead gen opportunity."

The plugin guidelines restrict most of that path. Plugins "may conduct commerce only for physical goods," and "Selling digital products or services, including subscriptions, digital content, tokens, or credits, is not allowed." Users may sign in with an existing paid account, but plugins cannot "display subscription plans, initiate new subscriptions, or promote upgrades." Checkout must happen on the developer's own domain, and embedded third-party checkout is prohibited. Plugins "must not serve advertisements and must not exist primarily as an advertising vehicle." Descriptions may not advertise pricing, promotions or trials.

Requiring an account to use the plugin is permitted. Running the plugin as a conversion surface for a SaaS subscription, which is what "convert the ones who look you up" means for most software businesses, runs into the no-subscription-promotion rule inside ChatGPT. The revenue has to be earned somewhere the plugin cannot mention.

"Ship five of them" is five domain verifications, twenty-five positive test cases and fifteen negative ones

Step 4 on screen reads "Ship 5 of them and see which ones get picked up," and the narrator repeats it at 0:57. OpenAI's submission documentation sets out what shipping one public plugin costs. The developer needs a verified identity, completed through individual or business verification in organization settings. Each MCP server needs domain verification, done by hosting a challenge token as plain text at .well-known/openai-apps-challenge on the verified domain. Each submission for review requires "Five positive test cases. Include the scenario, user prompt, expected tools, and expected result" and "Three negative test cases. Include a prompt or scenario where the plugin should not act, why it should not complete the request." Each one also needs a dedicated reviewer test account with a login URL, workspace details and sign-in instructions that work without multi-factor delays. The package ships as a ZIP containing the manifest, skills directory, icons and MCP configuration, goes through automated metadata checks, and then waits in a review queue with no published turnaround time.

Multiplied by five, that is five domain verifications, twenty-five positive test cases, fifteen negative test cases, five reviewer accounts and five review queues. The video's framing of a week of playing around does not account for any of it.

Two of the three platform analogies hold, and the TikTok one is off by a quarter

The post's historical anchors are checkable. Facebook Platform launched on May 24, 2007 at f8 with 65 developer partners and over 85 applications. Apple's App Store opened with more than 500 native applications the day before the iPhone 3G went on sale on July 11, 2008, with more than 125 of them free. Both match "Facebook apps in 2007" and "iPhone apps in 2008."

"TikTok in 2019" is the weak one. Quarterly worldwide download data puts TikTok at 15.2M, 18.4M, 24.6M and 75.4M across the four quarters of 2018, then 45.2M, 41.0M, 39.1M and 57.3M across 2019. The breakout quarter was Q4 2018, a 206% jump driven by roughly 60 million Musical.ly users migrating after ByteDance completed the rebrand in August 2018. Three of the four 2019 quarters came in below that Q4 2018 peak. Annual 2019 volume did exceed 2018, so the year is not wrong as a period of high growth, but the inflection the analogy is reaching for happened in late 2018. The deeper mismatch is category: Facebook Platform and the App Store were developer distribution platforms with submission flows, and TikTok's 2018 and 2019 window was organic consumer reach with no developer surface at all.

Key Takeaways

  • Correction: ChatGPT did not start recommending plugins mid-conversation on September 29, 2026. OpenAI's own developer forum shows plugins being "Suggested in conversations daily 09-04 to 09-09," with developers filing bug reports when that surface broke on September 9. DevDay improved ranking and recommendations on an existing surface.
  • Correction: Step 3 of the playbook, writing plugin descriptions with the exact phrases people type into ChatGPT, is prohibited by OpenAI's plugin guidelines, which ban model-readable fields "designed to manipulate how the model selects plugins" and warn against overly generic names.
  • Correction: The causal order on Pietro Schirano is reversed. MagicPath was live as a ChatGPT plugin on September 2, 2026, posted a sidebar integration "in collaboration with OpenAI" on DevDay itself, and posted the 2,000% chart the next day. The company had already raised a $6.6M seed led by Khosla Ventures and Abstract. This is a launch-partner outcome, not an empty-slot outcome.
  • Partial correction: "TikTok in 2019" misses the inflection. TikTok's breakout quarter was Q4 2018 at 75.4M downloads, and three of the four 2019 quarters landed below it. TikTok was also never a developer distribution platform, which makes it a poor third term next to Facebook Platform and the App Store.
  • Verified: The 1.2 billion weekly ChatGPT users figure is accurate and traceable to Sam Altman's DevDay keynote on September 29, 2026.
  • Verified: "Facebook apps in 2007" and "iPhone apps in 2008" are both correct, at May 24, 2007 and July 10 to 11, 2008 respectively.
  • Verified: The on-screen caption spells "Greg Isenberg" correctly. The Whisper transcript at transcript.txt line 37 renders it "Greg Eisenberg," which is wrong. Isenberg is CEO of Late Checkout and hosts The Startup Ideas Podcast.
  • Unstated cost: Each public plugin requires verified developer identity, domain verification via a .well-known/openai-apps-challenge token, five positive and three negative test cases, a working reviewer test account, and a review queue with no published turnaround. "Ship 5 of them" means doing all of that five times.
  • Unstated cost: The lead-gen monetization answer the clip endorses collides with OpenAI's rules. Plugins may transact only in physical goods, may not sell subscriptions or digital content, may not display subscription plans or promote upgrades, must check out on the developer's own domain, and may not serve advertisements.
  • Unverified: "Millions of questions get asked every day with no plugin to recommend" has no published source. OpenAI has not released a figure for unmatched queries, and the "empty slot" count that the whole thesis rests on is an assertion.
  • Unverified: Schirano's 2,000% has no baseline. The chart in frame 39 shows no y-axis label, no units and no starting value.

Resources

Published October 2, 2026. Writeup generated from a favorited TikTok.