32 growth playbooks, each with its number. A company's own pricing page is cited first in only 12% of AI answers.

Most growth writing gives advice. This is a record of what happened. Every entry below names the company or person, gives a number, and links to a source you can open. The thirty-two entries were collected between mid-August and this week. They come from founders' post-mortems, vendor disclosures, benchmark studies and one interview given three days ago.

Read them together and one changes how you read the rest.

The one insight to take away

Kyle Poyar and Nikolas Laskaris collected 7,600 answers from six AI engines to pricing questions about the Cloud 100 between 29 July and 10 August. The company's own pricing page was cited first in only 12% of answers.

Plaid's page came first 42% of the time, the best result in the set. Yet the page doing most of the work was its billing documentation. It appeared in 70% of answers, ahead of the pricing page at 64%.

The docs define one-time, subscription and per-request billing down to the endpoint. They cover the expensive edge cases and state which figures are not public (Poyar and Laskaris).

The same study found that 57 of 77 public Cloud 100 pricing pages were fully readable by a bot. Ten hid at least 40% of their body.

Profound, whose data the study used, had the same problem on its own site. Its prices rendered client-side in JavaScript, so a crawler fetching the raw HTML saw no prices. It moved them to server-side rendering on 25 June. Citation traffic rose 13% week over week, and the pricing page became the site's second most-cited page (Poyar and Laskaris).

Now look at where the citations went instead. Promptwatch measured Reddit's share of ChatGPT Search citations at 3.83% from 18 July to 7 August, one of the largest shares of any domain. From 14 to 17 August, it averaged 0.52%. That was an 86% drop in two steps (Promptwatch).

Otterly measured a fall of at least 73% across 16 brand reports. It found that official pages and reference sources filled the gap, not other forums (Otterly). Trellner found three sites on one Cloudflare nameserver pair that had published 215,128 generated "best software" pages since December 2023. Perplexity cites them (Trellner).

Put the four together. The page that wins is the one a machine can quote. A form that says "contact us" cannot be quoted. Neither can a price hidden behind a JavaScript render. Reddit threads used to be quoted for you. Now they are not. Dense documentation with real numbers in the initial HTML gets lifted.

The check takes a minute. Fetch your pricing page with curl, or open it with JavaScript off, and read what comes back. If the prices are missing, you are one of the ten.

We ran the same test this morning on TIME, following Vincent Schmalbach's method. We fetched the site and changed only the user agent. A browser gets the full page. ClaudeBot, PerplexityBot and GPTBot get text/markdown.

The headers include x-mobian-registry-version: 2026-09-11.v11. The text includes a block that starts "Sponsored content. Supplied in partnership with…". TIME is giving crawlers a cheaper page with the ad baked into the material the model ingests (Schmalbach, reproduced 13 Sep 2026). It is the publisher's version of the same insight.

Two samples, so you can see what the entries look like

Pocket FM took its run rate from about $250M to $500M in a year, and the lever was catalogue breadth, not cheaper content. AI now produces 99% of new content and covers 93% of the catalogue. Production is about 80× cheaper. Producing 100 hours used to take a year; now it takes a day.

That let 550,000 creators produce 2.5 million hours a year. Two years ago, the entire catalogue contained about 100,000 hours. CEO Rohan Nayak credits that breadth for 12-month revenue retention rising from 44% to 76%.

Out of 770,000 series, 96 titles have passed $1M each and 13 have passed $10M. "ARR" here is monthly revenue times twelve, and he said so. Profitability is "on an adjusted basis" with margins undisclosed (TechCrunch, 10 Sep 2026).

Dayzle spent $220 on Google app ads. 13 of 56 billed installs were people. Thirty-three installs used an app version the Play Store had stopped serving days earlier. Each was sideloaded from a saved file, opened once and used for zero seconds. Seven came from countries outside the targeting.

Here was the loop: the campaign goal was installs. A farm installed the app from a saved APK, Google counted the install, and each fake conversion taught the algorithm to send the farm more ads.

The fix was to change the goal to "won a puzzle", an in-app event that takes real effort for a script to fake. In Nick Abe's words, "the idea is just to make us more expensive to farm than the next app" (Dayzle).

Below the line are all 32, grouped into six families. Each entry records the number, the mechanism you can copy, and where the source is thinner than it looks.

How to use this library

Every entry has the same parts. The number appears in the source's own terms. The mechanism is the part you can run on Monday. Where it applies, the caveat explains why a self-reported figure and a measured one are not equivalent.

Every entry links to its source. If a figure is single-sourced or comes from a vendor measuring itself, the entry says so.

The families run from the newest problem to the oldest.

Family 1 — Getting quoted by the machine

1. Plaid's billing docs are cited more than its pricing page. Across 7,600 answers on six engines, Plaid was cited first in 42% of runs. Its billing docs appeared in 70% of answers, its pricing page in 64%, and its FAQs in 50%. Across the full set, the figures were 12% first and 46% anywhere.

The docs define billing at endpoint level and state which information is not public. That gives the engine an authoritative explanation for a missing figure instead of leaving it to guess from a forum.

The next-best result was Fireworks AI. It publishes scheduled price changes as dated columns and exposes an llms.txt index (Poyar and Laskaris). Caveat: the dataset is Profound's, a commercial partner of the publication. Treat the ranking as solid and any single percentage as single-sourced.

2. Profound moved pricing to server-side rendering; citations rose 13% in a week. It deployed the change on 25 June 2026. The same study found four causes of unreadable pricing pages: client-side rendering, content loaded only after a tab or accordion click, an old robots.txt block, and a pricing widget on an uncrawlable iframe domain (Poyar and Laskaris). Caveat: the result was self-reported by the vendor whose product ran the study. You can check the mechanism on your own site with one command.

3. Three pages get you quoted after the Reddit collapse. Tom Orbach's mechanism, paired with Promptwatch's number, starts with a pricing page containing real figures. A model cannot quote "contact us".

Next is a "[you] vs [competitor]" page. Buyers constantly ask for comparisons, and the engine quotes whoever has written one. The third page is an FAQ that answers each question in its first sentence, using customers' words.

Reddit still gets cited when a named employee answers in a large thread. Anonymous accounts do not carry through (Orbach; Promptwatch). Caveat: Promptwatch calls the size of the drop provisional, and OpenAI told reporters it sets no fixed visibility level for any site.

4. Three linked sites published 215,128 "best software" pages, and Perplexity cites them. Two web-grounded Perplexity models were asked for the best products in 380 buyer-intent categories. They produced 7,534 citations across 2,055 domains.

Of those citations, 59.8% pointed to domains ranked worse than 100,000 on Tranco. Another 23.4% pointed to domains outside the top million. Wikipedia received 3 citations.

The three sites were registered through the same registrar between December 2023 and May 2024. They delegate DNS to the same nameserver pair, use the same template, and each has a blog containing exactly six posts about the other two.

The same "project estimation software" category produces three different rankings, each in JSON-LD. One site sells custom market research "from €5,000" above the taxonomy the models retrieve (Trellner). Caveat: only Perplexity was measured. The report makes no claim about other engines.

5. TIME serves crawlers a different website. Humans receive 303,235 bytes of HTML. Selected AI crawlers receive 13,409 bytes of markdown containing sponsored content. The ad partner is Mobian, which appears in the response headers. TIME reports that bot traffic exceeds human traffic on most days (Schmalbach).

Our reproduction, 13 Sep 2026: crawler user agents receive text/markdown with x-mobian-* headers and a sponsored block. One detail has changed since August. GPTBot returned 406 in Schmalbach's test but now receives the markdown too.

Family 2 — When the buyer is an agent

6. 30% of Vercel's deployments were started by a coding agent, up 1000% in six months. That is Vercel's own statement from April 2026.

The study citing it ran 16,893 sessions across 1,163 prompt variations, 75 repositories and three agents (Claude Code, Codex, Cursor). The agent was implementing the choice of third-party service, not recommending it.

The buyer persona barely changed the outcome. Adding a human approval step did. In the object-storage test, Amazon S3 won by default. Once an approval prompt was added, Cloudflare R2 began winning those sessions (Armature). Caveat: Armature sells growth services to dev tools and says so. The full traces are published.

7. Lazyweb grew to 50,000+ agents with a measured result per tactic. Auto-routing to the MCP from agents.md produced +100% DAUs. Guest accounts, which let an agent self-serve without an email, lifted activation by +50%. A playground where humans could try it before connecting an agent increased MCP activation by +40%. A one-line install command added +20%.

Every successful tactic removed a step an agent cannot perform. A Million Dollar Homepage revival, Facebook group infiltration and Reddit failed. Replying in X threads under 10,000 views worked, converting at roughly 20% (Ali Abouelatta).

8. Bill the agent by the outcome, not the attempt. Intercom's Fin charges $0.99 per resolved conversation, with unresolved conversations free. From May, Zendesk's "Verified Resolution" is confirmed by an LLM evaluation within 72 hours. It costs roughly $1.20 to $1.50 on committed volume, while assisted escalations are free.

Salesforce's Agentforce launched at $2 per 24-hour session, resolved or not. Customers called that impossible to forecast.

The sequence is simple. Pick a unit that exists as a database field. Verify it on a clock, make near-misses free, and absorb the failures yourself. Under the old model, one developer running a hundred agents in parallel accumulated $1.3M in tokens in thirty days (The Next Web). Caveat: the outlet itself says the report that OpenAI is offering outcome pricing to select accounts is unconfirmed.

9. The seat model is dying of price increases, not of AI. The Vertice SaaS Inflation Index ran between 12% and 16.4% through 2026, compared with roughly 2.7% general inflation. It peaked at 14.7% in Q4 2025 during renewal season.

Zylo's 2026 index says the average enterprise spends $55.7M a year on SaaS, up 8%, while its portfolio remains flat at 305 applications. The test for whatever replaces seats fits into one sentence: can you name one countable thing your software does that a customer would pay for on its own (Jason Lemkin, SaaStr).

10. Reducto cut document parsing from 3–6 cents a page to one. r-1 costs 1 cent a page all in. There is no accuracy tier and no credit multiplier. In early preview, Reducto reported a 20% lower error rate than its own previous pipeline.

The change is about legibility as much as capability. One per-unit price can be quoted by a buyer, entered into a spreadsheet and, as entry 1 shows, picked up by an engine (Reducto). Caveat: this is the vendor's own blog. The error figure compares the product with its own pipeline, not an independent benchmark.

Family 3 — The first customers

11. Your first customer is not a stranger. Make three lists: contacts (aim for 50), communities you already belong to (10 to 20), and anyone who has ever paid you for anything. Star the ten people most likely to buy.

Then make ten asks. Each should contain the problem, the solution, a price, a date and a yes-or-no question. "I'm taking on 3 [type of people] this month who are dealing with [problem]. I'll [specific outcome] for $[price], done in [timeframe]."

Deliver the work by hand before building software. Repeat until you have three (Noah Kagan).

12. One channel per MRR band, in order. The company reached over $4M ARR in one year and was accepted by YC. According to the founder, there was no magic channel. He "experimentmaxxed one channel at a time, based on our MRR level".

The first band is explicit: $0 to $6K MRR came from pure cold outbound (Dylan Txa, GojiberryAI). Caveat: the bands above $6K were not captured. The sequencing is the transferable part.

13. He became the Japanese-learning creator a month before the app existed. Kiku launched 30 March 2026. Posting on @haks_room began 24 February and focused on the problem, not the app. One March video about Japanese slang received 1.3M Instagram views.

Next came a faceless product account (+700K views), followed by three ambassador creators. One ambassador reached 77K views and 3K comments using a comment-for-the-link hook.

After five months, gross sales were about $16,000. Net sales were $12,200, with $7,000 in the most recent month. The price was about $50 a year with a 7-day trial.

"Almost all just from me posting 4 or 5 times a day organically" (Hakeem Dimozantos, public build log on @haks_room; the app is at kikulang.com). Caveat: the founder reported the revenue figures in a public video. They were not checked against the App Store.

14. Four weeks to a working product, three more to launch, and the first ~300 users onboarded by hand. Roman Ugarte previously led growth at Cursor as it grew from 15 people to over 1,000.

He built Grok Bot from scratch rather than adding it to Cursor. His reasoning was that a knowledge-work agent inherits the wrong assumptions inside a code editor. Even with enough distribution to skip personal onboarding, the team chose to do it (Roman Ugarte). Caveat: the headline numbers appear on the free portion of the page. The rest is paywalled.

15. A free month, on one condition: cancel a competitor first. A Cursor employee posted from her own account. She offered free Cursor Ultra, the $200-a-month plan, to anyone who cancelled a competitor and sent her a screenshot. A second employee repeated the offer with a 24-hour deadline and got a better result.

The steps are specific. An employee posts, not the brand. Require proof, ideally the last invoice so you learn what they paid. Give away months of your most expensive plan, never cash. Deliver it over DM and ask "what made you cancel them?".

Before the month ends, offer to hold them at whatever the competitor charged. People post their cancellation screenshots themselves (Tom Orbach).

Family 4 — Paid, trials and free

16. Dayzle: $220, 56 billed installs, 13 people. The details are above (Dayzle). Check: reconcile Google's install count with your own analytics before trusting it. Never optimise paid installs to the install.

17. 7-day trials convert 10% better than 3-day, across 10,000+ apps. The data covers Superwall's entire platform and tens of millions of trials.

The reusable move is to split the measurement. Track paywall-view to trial-start separately from trial-start to paid, because a shorter trial can win the first step and lose the second. The 14-day figures come from a smaller sample, and the post says so (The Paywall Index). Caveat: the data belongs to the vendor and was published by the vendor.

18. What separates AI apps that retain from the ones that churn. RevenueCat studied subscription data from 3,519 AI apps. Year-one retention was 13.9% for the high group, 5.3% for the mid group and 1.4% for the low group.

The difference appears at the first renewal. At high-retention apps, 57.9% of monthly subscribers renew once, compared with 30.2% at low-retention apps. By month three, the gap narrows to 79.5% vs 68.5%.

A 7-day trial is 12.7 points more common among high retainers. Subscription-only monetisation is 16.4 points more common. AI apps earn 41% more per payer in year one ($30.16 vs $21.37) and churn at 30% higher rates (RevenueCat). Caveat: RevenueCat's own platform data, so apps outside it are not represented.

19. Lovable gives away inference and holds it to a three-month payback. Free usage is booked as acquisition spend, not a cost centre. It competes with every other channel for the same budget. If $X of free product goes out, the question is how quickly it returns through conversion, retention and expansion.

The internal argument starts with the alternative. Keeping the money is not the alternative to giving product away. Spending multiples more to acquire the same customers elsewhere is.

This question bites harder in AI than in SaaS. SaaS ran at 80 to 90% gross margin. AI products can run at zero, and 40% is described as doing well (Elena Verna).

20. Flip the frame your whole category ignores. Regulatory Focus Theory divides buyers into promotion, where they chase an upside, and prevention, where they guard against loss. January fitness marketing is all promotion.

Chipotle renamed Quitters Day, the second Friday in January when resolutions die, as No Quitters Day. It ran the campaign through Strava segment challenges. The result was 190,000+ athletes across six cities, 700M+ earned media impressions and a 20% increase in Lifestyle Bowl revenue.

The cheap version is to write the opposite-focus twin of your top headline and run both (Strava case study; the framework is Tory Higgins' Regulatory Focus Theory). Caveat: the flip is a campaign, not a voice. Use prevention constantly and the brand gets tagged with risk.

Family 5 — Audience businesses

21. 100,000 subscribers with no research phase. Tom Orbach grew from a 700-person waitlist in August 2023 to 100,000 subscribers and 2,000+ paid subscribers in three years.

Every idea goes into one note built over years. More than ten articles are in progress at once, so a dry week costs nothing. A $150-a-month intern forwards the campaigns she runs into.

His publish gate is a generalisation test. If only the original company could run the play, he drops it (Orbach). Caveat: the scout and the flow of inbound campaigns both depend on already being known. The drafting habits do not.

22. $1M in year one with zero employees, and the split behind it. In Kyle Poyar's first full year working solo, he reached 88,000+ subscribers, up about 15% on the year, and 700+ premium subscribers at $15 a month, from zero premium subscribers twelve months earlier. He also advised 15 companies.

The revenue mix is the playbook: 55% brand partnerships, 25% consulting and advising, 10 to 15% subscriptions, 5 to 10% speaking. Subscriptions are the smallest slice.

His three rules are to never charge per hour, raise prices early as a filter, and choose a few long-term partners over many one-offs. He itemised the friction too: 65 emails and 94 days to collect one $10K invoice, plus three legal entities in a year (Kyle Poyar).

23. A LinkedIn newsletter put ~24,000 subscribers on the board in 24 hours, and 100 to 500 into Kit per send. Vojtech Vosecky has ~185,000 LinkedIn followers. He republishes long-form pieces as LinkedIn newsletter issues and puts job listings inside them.

He gates the lead magnet behind an email signup on Kit. He also delays the republished version, making the LinkedIn copy a funnel instead of a substitute.

The same piece gives a counterexample. Jay Clouse built a LinkedIn newsletter with 40,000+ subscribers and abandoned it because the subscribers were not his and he had no click data. The same product produced opposite results. The difference was whether it converted readers to an owned list (Jay Clouse, reporting Vosecky). Caveat: the essay reports all figures. They were not checked against LinkedIn or Kit.

24. $101,666 from a 3-day free summit, on $23,300 of cost. The revenue stack was $63K from the cohort, $17K from the downsell, $11K from VIP tickets and $9K from the upsell. There were 3,145 registrations and 1,000+ people live on day one.

Costs included a 15% performance fee to the partner agency, $7,800 for the domain and $1,800 for affiliates. The biggest surprise was the $47 VIP ticket at a free event. It produced $11K and identified the buyers before the pitch.

The whole project began with a cold email Clouse ignored twice. The third email contained a Loom showing work already done (Jay Clouse). Caveat: the source rounds the margin to $78,400. The subtraction is $78,366. An earlier summary said "3,100+ attendees", while the source says registered.

25. A fully synthetic creator reached 1,300 followers on about $100 of credits. Olivia Moore created "Janie", a 19-year-old going through sorority rush at Alabama. She started with one ChatGPT image, animated it with MiniMax and Grok Imagine, and added a voice with ElevenLabs. The work took about 30 minutes a day.

Within a week, the account had 1,300 followers. Videos had tens of thousands of views, with the first approaching 100K. TikTok labelled 8 of 20 videos as AI with no visible hit. Viewers spotted the fake by day two and kept watching (Olivia Moore, a16z). Caveat: Moore says she omitted TikTok's AI disclosure to test detection and broke the rules. The reproducible part is the pipeline and cost.

Family 6 — Scale, and cutting

26. Pocket FM: catalogue breadth drives retention. The details are above (TechCrunch). The control case is Pocket Saga, a three-month-old, US-only, fully AI-produced microdrama app. It is running at about $15M.

Humans still originate the stories on Pocket FM. "We want to create great IPs that last 100 years, and that needs humans."

27. Base44: a solo founder, $1M ARR in three weeks, an $80M exit in six months. The company reached 400,000+ users without paid marketing or outside funding.

The mechanism was to split the product by layer, let agents own the front-end build, and use building in public as the distribution channel. The line worth keeping is that removing a feature tripled activation (Maor Shlomo, interviewed by Lenny Rachitsky).

The deal was worth up to $80M, contingent on revenue targets through 2029. Wix's Q4 2025 report shows Shlomo on track for an additional $90M in milestone payments, more than the headline price (Calcalist). Caveat: the podcast is from July 2025.

28. Lovable picked up $10M ARR in 60 days on a brand prototyped in 30 minutes. Nad Chishtie joined eight months before the November 2024 launch as the fourth employee and first designer. He shipped an identity that was not built to scale.

When the product went viral, he rebuilt the identity by asking every paying user he could reach what the product made them feel. The design team remained one person for the entire period, by choice (Nad Chishtie, First Round Review).

29. Cutting $113,094 of ARR on purpose. Jay Clouse commissioned churn interviews for The Lab, then removed two of its three tiers. The change cancelled 106 Basic subscriptions and downgraded 24 VIPs.

The Basic forum was quiet while the Standard forum was active, so Basic gave members a false preview of the product. VIP received only positive feedback. Clouse cut it anyway because it consumed the founder's coaching time.

Retiring members are supported through July 2027 (Jay Clouse). This is the same operator as entry 24, moving in the opposite direction.

30. De-bundle the thing that grew inside your new product and ship it into the old one. PostHog built PostHog Code around one question: "if a team powered by AI rebuilt your flagship product today, what would they build and what would they skip?"

The team watched which capability grew inside it, self-driving, and extracted that capability into the main product. The original vessel continued changing shape into PostHog Desktop. The new product is a discovery vehicle, not a fortress (Cleo Lant, PostHog). Caveat: this is a first-party account of work in progress. Treat it as a method, not a result.

31. The decline curve, published by the founder. Bank Statement Converter's revenue is down 24% and MRR is down 12% from the February 2026 peak. New subscribers fell from 191 a month in January to 45 in August. Cancellations stayed flat.

The diagnosis is the founder's own. Churn did not spike; acquisition collapsed because users with a few pages now do the conversion in a free chatbot.

There was a second-order effect. He stopped building in public after watching people clone the product from his revenue screenshots (Angus Cheng).

32. Eight years of building a creator business so it could be sold. Pat Flynn and Matt Gartland made Smart Passive Income a separate entity, removed Pat's face from the website, and made the community the hero. They prepared for a sale they were not sure would happen.

The deal had no broker. It used seller financing and required 160 documents for diligence. It closed in six months.

The buyer, Liz Wilcox, closed five months before the announcement. In the meantime, she publicly worked as "director of community". The community doubled during that window (Flynn and Gartland, interviewed by Jay Clouse). Caveat: the sale price was never disclosed. Do not infer one.

What is not in here

A funding round is not a playbook, so none are listed. Four entries contain numbers from vendors' own data (2, 10, 17, 18), and each says so. One entry from the source folder was omitted because its only source was an unbylined newsletter body. It will be added when a primary source turns up.

If you take one thing from the library, take entry 1. Then fetch your own pricing page and read what the machine reads.