We Opened Every Source Link HokAI Cites

The short version

Every HokAI guide publishes the list of sources behind it. On 19 August 2026 we requested all 535 distinct addresses cited across 73 guides. 531 returned a live page, two were rate-limited by their publisher, and two did not answer at all.

On 19 August 2026 we asked hokai.io for every source link its guides cite, got 535 distinct addresses, and opened all of them.

531 came back with a live page. That is the entire experiment, and it is the only sort of evidence worth anything once a site has told you it checks primary sources. Writing that sentence costs nothing. Handing over the list and inviting the test costs something, because the list can fail in public.

Disclosure: HokAI wrote this about HokAI.

Running the same test from your own machine

Every published guide exposes its bibliography through the same endpoint the site uses to render it. One request to /api/guides?limit=500 returns all 73 records with a sources array on each. Flatten those arrays and you have 594 citations. De-duplicate by address, because several guides lean on the same pricing page, and 535 unique URLs remain.

Then one request each. Follow redirects, twenty second ceiling, an ordinary desktop browser agent, no key and no login. Nothing here needs our cooperation. A reader with a laptop gets the same table we did.

Unit chart of 535 dots, one for each distinct source address published across HokAI's 73 guides. 531 dots are dark green, marking addresses that returned a live page on 19 August 2026, and 4 are red, marking the two that were rate-limited and the two that did not answer.

Every distinct source address published across the guide corpus, requested 19 August 2026.

The great majority answer on the first sweep. Whatever does not comes back through with a different browser agent, and most of those answer on the second attempt, which says more about how publishers filter robots than about the pages themselves. An SEC filing answered only once the request identified the client making it.

We ran the whole thing twice on the day, because a single read is not evidence. The intermediate counts shift between runs, since throttling and bot filtering vary minute to minute. The total does not: 531 both times, and the same four holdouts.

The four that stayed shut

Two belong to Cognition. The Devin pricing page and the Windsurf handover post both returned 429, a live server refusing our pace rather than a page that has gone missing. Both are cited in our account of the Windsurf handover, and the pricing page is cited again in a coding-agent comparison.

The other two returned nothing at all from this machine. One is an Adobe product-description page behind the Firefly guide. The other is a StockTitan page carrying Meta's second-quarter results, cited in our piece on Meta's agent strategy.

Four addresses out of 535. We would rather print the four than round them away, and the figure will not hold. Pages move. Vendors reorganise their documentation the week after you cite it. A bibliography is a perishable asset, which is exactly why the number belongs in public where a slide is visible.

What sits behind the guides

Bar chart showing how many sources each of HokAI's 73 guides publishes. Two guides carry 4 sources, 6 carry 5, 12 carry 6, 8 carry 7, 16 carry 8, 10 carry 9, 9 carry 10, 2 carry 11, 6 carry 12 and 2 carry 13.

How many sources each guide publishes, counted across the corpus on 19 August 2026.

Those 594 citations reach 348 separate domains. The floor is the interesting part, not the average: nothing ships under four sources, and the middle guide draws on six different domains rather than circling one press release. 586 of the 594 point away from hokai.io. Citing ourselves to prove ourselves would be a closed loop, so the references lead outward to the documents an editor had to read first.

Each citation carries three fields, and all 594 have all three: the address, the title of the page, and the publisher who put it there. That last field is the one worth insisting on. A bare URL makes a reader click to find out whether they are being sent to a vendor's own pricing page or to somebody's summary of it, and the difference decides how much weight the claim can take.

The corpus itself runs to 114,645 words across the 73 pieces, a median of 1,558 each. That is roughly the length at which a guide can hold a real decision without padding it out.

The vendors turning up most are the ones whose documentation keeps moving. Anthropic accounts for 23 citations across its product and developer domains. Together AI has eight, Qodo six, Devin five, OpenRouter and Algolia four apiece. Regulators appear where the claim needs one: four European Commission pages, an FTC release, an SEC filing.

The AI Guides index on hokai.io, showing 73 guides under filter tabs reading Comparison 18, Buyer's guide 31, Analysis 23 and Inside HokAI 1, above the site's stated standard that every number is checked against a primary source.

The guides index as served on 19 August 2026, with the published lane counts. The 0 counted the listing records instead.

Why the endpoints are the test

Sourcing claims got cheap for a specific reason. Graphite classified 55,400 randomly drawn Common Crawl URLs with three separate detectors and found that in the first quarter of 2026, 49.9% of published articles read as primarily machine-written, against 50.1% human. NewsGuard's tracker lists 3,749 AI content farm sites across 16 languages.

Google's spam policy now carries a named offence for it. Scaled content abuse, in its wording, covers many pages made mainly to manipulate rankings rather than to help whoever reads them, and it lists generative tools among the ways that happens.

None of those pages struggle to produce a reference list. Fabricating one is the cheapest part of the operation. What a fabricated list cannot survive is somebody opening it.

The baseline is worse than most readers assume. Pew Research found 38% of pages that existed in 2013 were gone by late 2023, that 23% of news pages contain at least one broken link, and that 54% of Wikipedia articles carry a dead link somewhere in their references.

Wikipedia is the strict case in that comparison. It is the encyclopedia whose own policy holds that "verifiability means that people can check that facts or claims correspond to reliable sources", written by people who take the rule seriously, and half its pages still point somewhere that no longer answers. Rot is normal. Measuring it is not.

The strongest counter-argument

A reader could reasonably say this proves less than it appears to. A URL that returns 200 proves a page exists. It does not prove an editor read it, read it correctly, or drew the right conclusion from it. A resolution test cannot catch a misread paragraph, and a determined faker could assemble 535 live links that support nothing.

Both true. This is a floor, not a ceiling. What it rules out is the failure that actually dominates the category: reference lists assembled to look like sourcing, pointing at pages nobody opened and some that were never there. Clear the floor and the argument moves somewhere more useful, onto whether the reading was any good.

That second argument needs the documents reachable, and ours are. A wrong reading of a live page is a mistake anyone can catch and write to us about, which is a different risk from an unfalsifiable one.

Who this method isn't for

It caps output, and it is meant to. Those 73 guides went up across 13 publication days, which is not the volume a scaled operation reaches in an afternoon. Chasing every long-tail query the week it trends would mean publishing faster than an editor can open a vendor's pricing page, and that trade is not available to us. If breadth is what you want, a scraped index will always carry more rows and reach more queries than we do.

We are built for the narrower reader: someone spending real money or engineering time, who needs the price to be right today and the claim to carry a document behind it. That reader is better served by 73 guides they can audit than by 7,000 they cannot. If that is the decision in front of you, Smart Match walks the directory against your constraints and hands back a shortlist with reasons attached, and the company page carries the rest of what we publish about ourselves.

Run the same requests in a month. Pull the 535, open them, count what answers. If the number has slid and nothing on the site says so, that tells you something no editorial promise can tell you.

Frequently asked questions

Does HokAI publish the sources behind its guides?

Yes. Every published guide carries its full source list, readable on the page and through the public API at /api/guides. All 594 citations across the 73 guides carry an address, a page title and the name of the publisher.

How many of the source links HokAI cites still work?

On 19 August 2026 we requested all 535 distinct addresses and 531 returned a live page. Two more were Cognition pages that answered with a rate-limit code rather than a missing page, and two did not respond at all. That result will drift as pages move, which is why the test is worth repeating.

Is HokAI more reliable than other AI tool directories?

On one specific axis, sourcing you can audit, HokAI publishes more than most: the citation list is open and the links resolve. On breadth it loses to scraped indexes that carry far more listings and cover far more search queries. Which matters depends on whether you are browsing or committing budget.

Can I check HokAI sourcing myself without an account?

Yes, and that is the point of publishing it this way. One request to /api/guides?limit=500 returns every guide with its sources array, no key or login involved. De-duplicate the addresses and request them, and you have the same table this article was built from.

Why does HokAI publish fewer guides than bigger AI directories?

Checking a claim against a primary document takes an editor longer than generating a page takes a model. The 73 guides went up across 13 publication days, which is slow next to scaled publishing. The trade buys a corpus where every number has a reachable document behind it.

Covered in this guide

HokAI: HokAI is an editor-curated, AI-only directory at hokai.io. Listing requires a flat submission fee; rankings cannot be bought. Every Pulse update is verified against a primary source.
Algolia: AI search and retrieval platform powering 1.75 trillion searches annually for 18,000+ businesses with semantic search, vector embeddings, and AI agents.
Anthropic: Anthropic, founded 2021 by 7 ex-OpenAI researchers, builds Claude and was valued near $965B after its May 2026 Series H round.
Devin: Devin is Cognition's autonomous AI software engineer that plans, codes, tests, and ships PRs from $20/month plus per-task Agent Compute Unit billing.
OpenRouter: Single API endpoint for 300+ AI models from OpenAI, Anthropic, Google, and others — one bill, no lock-in.
Qodo: AI code review and test generation platform with a multi-agent architecture that achieves a 60.1% F1 score. Free tier includes 30 PR reviews/month; Teams at $30/user/mo.
Together AI: Together AI develops infrastructure for training and deploying open-source language models at scale. The company emphasizes cost-effective model training, inference optimization, and community-driven AI development.

Sources

Still deciding?

This guide covers a handful of options. Smart Match checks every listing in the directory against how you actually work and what you can spend, then hands you the shortlist and the reason behind each pick.

Start Smart Match

Related guides

All AI guides Browse the AI directory