How Speechify Work uses competitor funnels to kill product hypotheses before writing code

How we run it: three buckets before a hypothesis reaches the roadmap

Our internal numbers stay internal, so this case study only uses what is observable in HookRadar. The process itself is simple and repeatable. Every product or positioning hypothesis that looks promising goes through the same check against competitor funnels before anyone opens Figma.

"We stopped treating a competitor's ad as evidence that an idea works. The ad is a hypothesis. The landing page they are still paying for three months later is the result."

Pavel Gusak, Creative Producer, Speechify Work

Below is what that check looked like for our own category in September 2026, the AI-agents-for-work market, where our four closest competitors are Perplexity Computer, Manus, Lindy and sintra.ai.

1. AI work agents: four brands, four different funnels hiding inside one category

On the surface, Perplexity Computer, Manus, Lindy and sintra.ai sell the same promise: an AI that does the work instead of answering questions about it. One click past the ad, they are running four unrelated businesses.

Perplexity treats the landing page as the unit of testing. sintra treats the quiz as the unit of qualification. Manus treats the store listing as the whole funnel. Lindy tried to be two brands and went back to one.

The ad creatives alone would not tell you any of this. Perplexity's video hooks ("Here's how I saved $30,000 using AI", "19 AI models running at the same time") look like a brand campaign. The destination data shows a factory: 203 separate pages, each promising one finished deliverable for one trade or job.

2. One strong direction: Perplexity turns every trade into its own landing page, then kills 92% of them

Perplexity has the cleanest funnel experiment in the dataset: name the trade, name the document, build a page for it. Since the first launch spike on April 17 (732 ads in one day), the brand has shipped 203 distinct use-case pages under one path. Twenty-nine of them sell an estimate or bid for a trade (roofing, HVAC, plumbing, fencing, concrete, painting, pool service). Thirty-eight sell an audit of a small business (law firm, yoga studio, vet practice, CPA firm, nail salon).

The page is the hypothesis. Most pages got the same treatment: 95 of the 203 received exactly 30 to 49 ads, which reads as a templated test budget, not a bet. Only 9 pages ever crossed 100 ads.

Then the cull. As of September 21, live ads point to 17 of the 203 pages.

Perplexity built 203 use-case landing pages since April 17. On September 21, ads still point to 17 of them.

Each page sells one finished deliverable for one trade or job (roofing estimate, CPA firm audit, job applications). Of 203 pages, 17 (8.4%) are still funded and 186 were tested and left. The biggest bets that still died include api-platform (163 ads), seo, video-to-shorts and recruiting.

Perplexity's landing-page factory after five months. The survivors cluster in two verticals (construction, real estate) and three broad jobs (job search, SEO and marketing, finding money).

Perplexity's construction page is the one trade funnel that graduated from a 30-ad test to a 42-day video. The ad is not about AI. It is about a spreadsheet the estimator no longer has to build.

Two more funnels ran alongside and went nowhere. Forty-four creator landing pages (pplx.ai/<creator>, perplexity.ai/influencers/<name>) collected 72 ads between them, one to four each, and none is live. Three Zoom event pages for a September 3 live session ("Your warehouse has the answers, the company doesn't know SQL") carried 7 ads and were the only event funnel in the category this year.

The lesson is not that vertical pages fail. The lesson is that the brand with the most money in the category tested roughly 60 verticals and kept two. Two hypotheses on our own list were versions of pages sitting in that 92%. They did not get a brief.

3. Another strong direction: sintra sells "AI employees" and lets a quiz do the routing

sintra.ai runs the highest creative churn in the set (392 ads launched on September 3 alone) and the most legible funnel architecture. Every destination falls into one of six buckets, and the September 21 snapshot shows which buckets the brand still believes in.

sintra's live set: one owner-burnout statement, seven destinations. The ad does not change; the page does.

The quiz survived by naming the model you already pay for. Plain /quiz is dead. /quiz-gpt and /quiz-claude, launched the same day in August with 17 video variants between them, are both live. The current hook does not mention sintra's product at all.

Roles beat industries. Every industry page (restaurants, salons, clothing brands, jewelry, mechanics) has gone dark. Every role page is live, fed by one image family. The same copy is pointed at six different role pages plus the homepage, which means the destination is the variable being tested, not the creative.

What sintra has already killed, so we did not have to: every discount and price-anchor hook ("Up to 70% OFF", "PAY FOR 1, GET 12", "From $5,000+/mo to $15.60/mo", "Replace your $5k marketing agency in 5 minutes"), the head-to-head "ChatGPT vs Sintra" comparison (launched and stopped three times: July 27, August 10, August 31), long-form story images, and the cartoon "The team you could never afford. Hired." The brand that tests more than anyone in the category has stopped paying for all of them.

4. Manus and Lindy: install-first, and the funnel that never left the homepage

Manus has one funnel and refuses to test another. Of 2,590 tracked ads, 2,544 send to the App Store or Google Play. The website got 43 ads. A co-branded page with Nothing phones got 3. All 19 live ads are text-less vertical videos with a localized install button ("Install now", "Installer", "Cài đặt ngay") that have been running 50 to 68 days.

The creative does the qualifying that sintra's quiz does. HookRadar's Hook Insights rank Manus's Arabic transformation hook ("I thought preparing presentations took hours… until I tried this") at the 99th percentile of the whole category, and a Chinese "two months until my birthday, I need a set of stunning photos" hook at the 98th. Neither mentions agents, models or work. They sell one finished artifact to one person in their own language, then hand them to the store.

Manus's September wave: 300+ ads in ten days (September 4 to 14), every one of them to a store listing, then silence from September 15. The brand spends in bursts, not in tests.

Lindy tried to be two brands and came back as one. Lindy's 346 ads split across lindy.ai (177), lindy.ai/go (131) and a separate identity at meetingsai.com with four sub-pages (/brain, /never-blocked, /nobusywork, /doer). The meetingsai experiment absorbed 38 ads and has zero live. After going silent from July 17 to August 19, Lindy relaunched on September 11 and 17 with 41 ads, all to lindy.ai.

Lindy's relaunch drops the second brand, the Slack-inbox POV and the coworker-curiosity hooks. What is left is the promise every rival now makes: finished work, not a draft.

5. The borrowing map: which funnel ideas travel between brands, and which die on arrival

HookRadar tracks 15 brands across four adjacent categories for Speechify (AI work agents, reading and TTS apps, dictation, language learning), roughly 75,000 Meta ads and 670 destination URLs in total. That is enough to see the same funnel idea appear in several places and to check what happened to it each time.

Two patterns hold across the map. First, an idea that works in consumer subscriptions (quiz, second domain, creator page) does not port to AI work tools by default; only the version that names the buyer's existing tool or role made it. Second, when four brands converge on a funnel shape, the survivors are the narrowest versions of it. "AI for real estate agents" lives. "AI for restaurants" is dead at two brands.

6. Creative borrowing with dates: who picked up what, and how fast

Funnels move slowly. Hooks move in weeks. Ten weekly HookRadar reports (June 29 to September 7) plus the September 21 snapshot show the same angles passing between brands.

The co-worker hook is the fastest borrow on record here: Lindy's June original was dead by the time Manus and Perplexity shipped their versions in mid-August. By the time a pattern looks new to an outside team, it has usually finished one full cycle.

7. What we would tell another product team

A competitor's landing page is the cheapest user research you will ever run. They paid for the test. The only cost is looking one click past the ad.

Methodology and data notes

Sources. All competitor data is Meta Ad Library data as processed by HookRadar for the Speechify Work workspace: per-competitor destination summaries (180-day window ending September 21, 2026), the active-ad snapshot of September 21, ten weekly HookRadar AI Reports (June 29 to September 7), three brief packs (weeks of August 17, 24 and 31) and Hook Insights over 90 days. Cross-category comparisons use the same destination summaries for Speechify's other HookRadar workspaces (reading and TTS, dictation, language learning).

Coverage. 18,183 ads across the four AI-agent brands, of which 2,839 have HookRadar AI analysis. Across all 15 tracked brands: roughly 75,000 ads, 10,200 analysed, 670 distinct destination URLs.

Definitions. A "funnel" or "destination" is a distinct cleaned landing URL (query strings and /signin suffixes merged). "Live" or "still funded" means at least one ad pointing to that URL was active on September 21, 2026. Not live does not prove the page failed; it proves the brand stopped buying traffic to it in the observed window. Counts are numbers of ads, not spend. Impressions are EU-only, as reported by the Ad Library. HookRadar signal is a 0 to 10,000 creative-strength score, not an impression count. Thumbnails are HookRadar previews of the first frame or the static image.

Known artifacts. Perplexity's destination list includes 8,962 ads to fb.me and 822 to instagram.com; an identical 8,962 fb.me count appears under an unrelated brand in another workspace, so these were excluded from funnel counts. Manus install videos carry no text, so their hooks come from HookRadar's AI descriptions. Hook Insights for this workspace are flagged low_sample for the 90-day window and were used for ranking examples only, not for counts. Lindy had no active ads between July 17 and August 19, so its earlier hooks are dated from weekly reports rather than the snapshot.