In December 2025 my AI tools directory, AI Platforms List, had two jobs that asked DeepSeek for new tools and saved whatever came back into platforms.json. On December 9 I pushed a commit called "CRITICAL FIX: Stop DeepSeek from hallucinating fake platforms". The main change was this block in the prompt:
🚨 CRITICAL - DO NOT HALLUCINATE OR MAKE UP PLATFORMS! 🚨 - ONLY suggest platforms you can VERIFY exist - ONLY include platforms from REAL sources (Product Hunt, GitHub, news articles) - If you can't find ${platformsPerBatch} REAL platforms, return FEWER - Fake platforms will be automatically detected and removed - We prefer QUALITY over QUANTITY - 10 real platforms > 25 fake ones Each call went to deepseek-chat over the plain chat completions API. It couldn't search or open a page, so it had no way to verify anything.
In November 2026 I rebuilt the directory, and we opened every website the old list pointed to. Of the sites that only my two DeepSeek jobs had added, 41% were gone. For the tools the list started with, it was 5.7%.
Two jobs, one model
The daily job was a GitHub Action. It ran at 06:00 UTC from December 5, 2025 to January 11, 2026, in three rounds of up to 8 tools. Its prompt asked for tools that "Have a working website/product (no vaporware)" and ranked affiliate programs above everything else: "PRIORITIZE platforms with active affiliate/referral programs (highest priority!)". Every tool it added was saved with added_by: "ai-auto-discovery" and verified: false.
The mass job went in at 04:00 UTC on December 9, in a commit titled "Add ULTRA HIGH VOLUME discovery modes - 90,000 platforms/month". It ran inside a generator script that looped with 30 seconds between cycles. Every second cycle it called DeepSeek, and after each cycle it pushed its changes to GitHub. Between December 9 and December 18 it added 317 entries.
Both compared each new name and domain against the list to skip duplicates. Neither one opened the site before saving it.
A validator that let parked domains through
About two hours later I added a URL validator, and the generator ran it every tenth cycle. It sent a HEAD request (a GET if the site refused HEAD) with a 10-second timeout. To catch parked domains, it looked for eight strings in the final URL after redirects, such as sedo.com/search and hugedomains.com. It never read the page.
Its first run, at 06:12 UTC on December 9, checked 729 entries and flagged 96 for removal: 81 dead and 15 with DNS errors. Three minutes later I put 34 of them back, because they were real sites that refuse HEAD requests. That run flagged 0 parked domains.
The prompt fix came about an hour and a half later, and its commit message shows what the validator was up against: "63% of discovered platforms were FAKE (90 out of 143 in Cycle 71)" and "DeepSeek was making up platform names and guessing URLs".
The pages built on top
A template turned the list into pages. On January 8, 2026, a commit titled "Generate 1,140 vibe-coding content pages" added 301 comparison files, and every one of them opens with the same line: "Both are powerful vibe coding platforms, each with unique strengths and use cases." Zapier, an automation tool, got that line too. 26 of the files have undefined in the name, where a slug was missing. zapier-vs-undefined.json is titled "Zapier vs Claude for VS Code: Which is Better for Vibe Coding in 2026?"
Three and a half hours later came a second commit, "Add 3,023 comparison pages for all platform combinations". The next day another commit put the whole site at 4,836 pages and 2.75 million words.
What Google did in January
Search Console for the domain:
| Month | Impressions | Clicks |
|---|---|---|
| December 2025 | 994 | 5 |
| January 2026 | 9,840 | 57 |
| February 2026 | 2,357 | 0 |
| March 2026 | 945 | 0 |
| August 2026 | 45 | 0 |
The best day was January 19, with 1,265 impressions. January 24 had 502, and January 25 had 20. Google's Search Status Dashboard shows no ranking update in that stretch. The December 2025 core update ended on December 29, and the February 2026 Discover update started on February 5. The mass job's last entry was on December 18, and I turned the daily job off on January 11. I don't know which pages Google reacted to.
Opening every link, 11 months later
On November 8, 2026, we ran a crawler over all 1,031 websites in the old list (1,095 entries, since some tools shared a site). It made one request per site plus redirects, waited at most 15 seconds, never sent two requests at once to the same host, and identified itself as AIPlatformsListBot with a link back. Each site that failed got a second try a few minutes later and a third about half an hour after the first. We only counted a shutdown when the site itself said so. For each gone site, we asked the Internet Archive for a capture from 2023 to 2025 and opened it to tell a real page from a parked domain.
44 sites blocked automated checks. Of the other 987, 763 still work (68 of them under a new name or address) and 176 are gone, which is 17.8%. A parked or for-sale page was the most common way a gone site failed, 51 of the 176.
Split by who added each tool
| Added by | Sites checked | Gone | Share gone |
|---|---|---|---|
| The list I started with (entries marked verified) | 565 | 32 | 5.7% |
| Research batches, December and January | 76 | 9 | 11.8% |
| Daily DeepSeek job | 50 | 13 | 26.0% |
| Mass DeepSeek job, before the Dec 9 fix | 63 | 41 | 65.1% |
| Mass DeepSeek job, after it | 213 | 80 | 37.6% |
Another 20 checked sites had entries from more than one source, and 1 of those is gone. Counting the two DeepSeek jobs together, including one site both of them added, 134 of 327 checked sites are gone, or 41%.
After the Dec 9 commit, the mass job's gone share fell from 65.1% to 37.6%. That is still more than six times the rate of the list I started with. For 81 of the mass job's 121 gone sites, the Internet Archive has nothing real from 2023 to 2025: 42 have no capture at all, 37 only a parked page and 2 an empty one. Many of those 81 may never have been products.
The groups aren't the same age. Most of the original list was in place by June 2025, while the DeepSeek entries came in December, so they had different amounts of time to fail.
Reading the page, not the URL
A parked domain answers 200, so a status check passes it. The old validator only looked at the URL (trimmed):
// December 2025 const parkedIndicators = ['domain-for-sale', 'parked-domain', 'this-domain-is-for-sale', 'godaddy.com/parking', 'afternic.com', 'sedo.com/search', 'hugedomains.com', 'dan.com']; const isParked = parkedIndicators.some(i => finalUrl.toLowerCase().includes(i)); The check the list uses now reads the first 20,000 characters of the page (trimmed):
// November 2026 const PARKED = /(\b(is|may be) for sale\b|buy this domain|this domain is parked|domain parking|parked free|hugedomains|afternic|sedo\.com|dan\.com\/buy|request pricing from name\.ai|window\.location\.href\s*=\s*["']\/lander)/i; if (status < 200 || status >= 300) return { ok: false, reason: `http ${status}` }; if (PARKED.test(html.slice(0, 20_000)) || new URL(finalUrl).pathname.startsWith('/lander')) { return { ok: false, reason: 'parked' }; } Some parking pages bounce you to /lander with a line of JavaScript, which the last test catches. A site that now opens a different domain fails too. And since anyone can submit a URL, including one that points into your own network, the check follows redirects one hop at a time and refuses private addresses at every hop.
What changed on the list
When I relaunched the list in November 2026, I dropped the mass job's entries and the research batches. 21 sites that only the daily job had added are still listed, all real products (Krisp and Vectara among them), each opened again on November 10. A tool goes live only after the check above passes, and a person approves every free listing.
A mistake I found while writing this
The study I published on November 8 told bot entries apart by a discovered_by field and an id prefix. The daily job set neither, so its 55 entries were counted with the original list. With them moved, the original list's gone share drops from 7.3% to 5.7% and the bot side's from 36.4% to 35.0%, so the gap grows from five times to six. The study page now shows the corrected numbers.
The study, with the full CSV (free to reuse under CC BY 4.0): https://aiplatformslist.com/study/ai-tool-shutdowns-2026
I wrote this with Claude. It dug through my old repo and ran the checks behind every number before drafting the post. I read and edited it before publishing.
