How we verify

Cards are written and maintained by EffectHub Research — our own editorial process, run by the people who build this catalog. There is no personal byline yet, and inventing one would be worse than saying so: the accountable party is the site, and the method below is what that accountability is made of.

Every description is written by us from the vendor’s own material. We take facts from other catalogs — that a product exists, its canonical URL, sometimes a category hint — and never their copy. That is also why the catalog grows in waves: a copied paragraph is faster and worth nothing.

What EffectHub Research publishes as data — the datasets, the monthly reports and how to cite them — is collected on /research.

The vendor’s own site first: product pages, pricing page, docs, changelog. Around that sit open datasets and public lists, which tell us a tool exists in the first place and let us see how many independent sources carry it.

A field we cannot confirm stays empty. An empty field on a card is a deliberate answer — it means nobody here has verified it — and it is always better than a confident guess. Pricing in particular is read off the vendor’s own page rather than inferred from a tier name.

Counters are literal. Views, click-outs, saves, ratings and category totals are whatever the database returns. Nothing is rounded up, padded, or displayed as an aggregate we did not measure.

The figures at the top of task, profession, category, country, pricing and platform pages are counted from the live catalog, not typed in.

• Tracked tools are published listings whose status is not dead or acquired. A category page counts what its list shows, so it also leaves out tools whose last check failed (Possibly discontinued).
• Free tier means the tool’s pricing model is free or freemium.
• Paid plans are the cheapest published paid plan of each tool, in US dollars, counted only when it is billed monthly. Yearly, one-time, per-seat, per-minute, per-image, per-credit and per-request prices are left out, so “a month” is always a monthly price. A price is shown only when at least 3 tools publish one, and only from $1 a month up — below that, a “monthly” price in our data is a per-unit rate. Shares are given as percentages only from 20 tools; below that they read “9 of the 10”.
• Public API means we have confirmed a public API for the tool; “not confirmed” is not “none”.
• The date (“Figures as of …”) is the day the figures were computed. Prices come from each tool’s card, researched from the vendor’s pricing page and logged when they change; they are not re-read from every vendor on that day.
• A page shows these figures only once it has enough tools to count: 8 for a task, 12 for a profession, 10 for a category, 10 for a country, 12 for a pricing page, 10 for a pricing page for one category, and 12 for a platform.
• The figures are free to reuse under CC BY 4.0 with a link to the page they are on. Everything EffectHub publishes as data is collected on /research.

A card’s status and its Last verified date come from a check we run against the product itself, never from a claim on the vendor’s homepage. The checker requests the tool’s own address — a HEAD request first, a size-limited GET when the server will not answer HEAD — follows its redirects, and sorts what comes back into one of three answers:

• FailedThe domain no longer resolves, the server refuses or drops the connection or does not answer in time, the certificate is invalid, the address redirects in a loop, or the server answers 404, 410 or a 5xx error.
• Kept outA 401, 403, 429 or other 4xx answer, or a bot-challenge page such as Cloudflare's. That is a fact about our request, not about the product, so it never counts as a failure.
• ReachedAny other 2xx or 3xx answer — including one that ends on a different domain. The move is recorded on the tool's record for an editor to review, but not counted as a failure: a redirect is as often a rebrand as a sale.

What the checker cannot see. It reads the start of a page only to recognise a bot wall. A parked domain, a for-sale page or a site that now sells something else answers like any live site, so the checker keeps such a card Active until an editor looks. Those calls are made by hand, and no check overwrites them: an acquisition, and a card an editor has marked Discontinued.

One failed check is not a verdict — sites have outages and hosts throttle bots. The first failure marks a card Possibly discontinued; a second failure in a row, on a later run, marks it Discontinued; any check that reaches the site puts it back to Active, unless the verdict was an editor’s. Before a run writes anything, it checks a short list of reference sites that are always up: if too many of those fail, the problem is our own network, and the run stops without changing a single card.

Re-checks run on a rolling schedule — by default every three days for the tools people open and click most, weekly for the middle, and every 30 days for the long tail. The checker runs every hour and takes up to 200 cards a run, the one checked longest ago first; a card whose certificate on record has expired since its last check goes back into the queue at once. The Last verified date is the date of the most recent check, whatever it found, so on a long-tail card a date a few weeks old is the schedule working, not a page we forgot. A date older than its schedule means runs were missed, and the card does not pretend otherwise: a lapsed certificate nobody has looked at for more than 7 days is shown with the date we last saw it, never called expired.

• ActiveThe last check reached the product, or was kept out by a bot wall — a blocked request, not a sign the product is gone, so the card stays in category pages. What a bot wall does cost is the standing signal of the effect score: we could not look, so it is left unscored rather than scored zero.
• Possibly discontinuedThe last check failed. The card stays up with the warning attached and scores lower until a check reaches the site again; category pages leave it out unless the filter is set to Any status.
• DiscontinuedTwo failed checks in a row, on separate runs — or an editor's call, such as an acquisition or a parked domain, which no later check reverses. The card keeps its page and the reason we logged instead of being erased.

A card the automated checker has not reached yet reads Not yet verified instead of borrowing a date from somewhere else; its status is the one established during import research, and never-checked cards go to the front of the checker’s queue. So the field tells you which of the two you are looking at, rather than implying the whole catalog was re-tested this morning.

A tool confirmed gone keeps its page, with the reason we logged and the date of its last check. “Is this thing still alive?” is a real question, and a deleted page is not an answer to it.

Importing a tool does not publish it. A record becomes a page only once it clears a completeness gate, and until then it stays in our import pool as a draft with no URL at all — an unfinished card is a 404 here, not a thin page carrying a noindex. That is the whole defence against the hundred-thousand-page directory nobody can use.

First, four hard conditions. A miss on any of them means the card does not publish whatever else it has:

• The tool is alive at the last check. A bot wall counts as alive — it is a blocked request, not a death certificate.
• It has a tagline and a description written by us, not a scraped stub. The description floor is 120 characters, which is the thinnest real card the catalog has shipped.
• It sits in at least one category.
• At least three of the facts below are confirmed. A card that clears the bar on two heavy facts alone is a pricing table, not a listing.

Then the fact set. 11 facts, weighted by how much each one changes what a reader can actually do with the card. The weights below total 78; a card’s score is that total expressed out of 100, and 55 is the line between a draft and a page.

• 12How the tool makes money — free, freemium, paid, or trial-led.
• 10A concrete entry price, or confirmation that there is none.
• 12Three or more capabilities we could confirm against the vendor's own material.
• 8At least one answered question on the card.
• 6Two or more platforms or integrations, an API counting as one.
• 6The company behind the tool is named.
• 6Three or more alternatives named.
• 6Two or more independent sources, or two evidence entries behind the claims we make.
• 4The year the tool launched.
• 4At least one visitor review.
• 4A vendor offer on the tool — paused while the catalog is ours alone, so no card earns this one today.

Three of those facts — a question, a review, a deal — are earned by people rather than by research, and they stay in the scale on purpose: a card carrying them genuinely is more complete than one that does not. Vendor deals are switched off entirely at the moment, so that last weight is currently unearnable by anyone; it keeps its place rather than being redistributed, because moving weight around would quietly move the publish line for every tool in the catalog. The weights are sized so a fully researched card clears 55 without any of the three.

Every listing carries one number out of 100. It is a derivation, not a column: computed from the catalog’s own data every time a page renders, stored nowhere, editable by nobody. A stored score starts drifting from the data it claims to summarise the day it is written, and the first person to notice is the tool owner reading their own numbers.

Four signals, ordered by how hard each one is to fake:

• 35%Real demand — Outbound clicks to the tool, counted twice, plus page views on its listing. Attention we can observe rather than attention someone claims — weighed by how much of it there is, so a few dozen visits nudge the score and thousands carry the full share.
• 30%Reception — Visitor ratings, pulled toward the catalog average until enough of them land. Three glowing reviews move a score far less than thirty.
• 20%Corroboration — How many of our source catalogs list the tool independently. One directory is a listing; six is a track record.
• 15%Standing — Alive at the last check, and a listing complete enough to act on — priced, described, placed. A dead tool scores nothing here whatever else it has. A site that turns our checker away is not scored here at all: we could not look, so the signal is missing rather than zero.

A signal with no data is not scored zero — it is not scored. Its weight is redistributed across the signals that do have data, and the result is still out of 100. Without that rule a tool published last week would be capped at 35 for having no traffic history yet, which says something false about it.

A trace of data counts as no data for the same reason: eleven page views are noise, and scoring them would rank a tool below an identical one nobody has visited. So each signal has a floor it must clear before it counts at all, and past the floor a small sample counts in proportion to its size: demand carries half its share at a thousand interactions and nearly all of it well past that, the way a few ratings are pulled toward the catalog average before they decide anything. Listings print how many of the four they were measured on, which is what stops a high score off two signals passing itself off as a high score off four.

The score is a summary of what we have measured. It is not a review, not a recommendation, and not a claim that a tool with 80 is better for you than one with 60.

The effect score above answers how is this tool doing. The trust score answers a different question — what do we actually know about the site you are about to sign up for — and the two come apart constantly. A tool with heavy traffic and a five-star average whose URL answers nothing, that names no company and publishes no price, is a high effect score and a low trust score. Merging them into one figure is precisely how a reader ends up on that site.

Same construction as the effect score: derived at render time, stored nowhere, nobody can edit it. Eight checks, and the dial on every tool page prints the findings behind them — the score and the lines under it come out of one pass, so a page cannot show a green line for a check the arithmetic marked down.

Two of the checks read things no catalog field could tell us, and both are collected rather than asserted. The domain’s registration date comes from the registry that is authoritative for it, over RDAP — the protocol that replaced WHOIS — and it is the hardest signal here to fake: a template, a logo and a wall of reviews can be bought in an afternoon; a domain registered eight years ago cannot. The TLS certificate — who issued it, and whether it has expired — comes off the handshake our own liveness crawler was already completing, at no extra request to the site. It is a sighting with a date: a handshake never returns a certificate that has already lapsed, so “expired” is always today’s date against one we read earlier, and we only say it off an observation from the past 7 days — a sighting of that certificate, or a check refused over it.

• 18%Site is up — Our crawler probes the tool's own URL on a schedule. A site that stopped answering, or that answers only to say the page is gone, loses this outright — and the date of the last probe is on the card.
• 16%Domain age — How long the domain itself has been registered, read from the registry over RDAP. A scam site can buy a template, a logo and reviews in an afternoon; it cannot buy a domain that was registered eight years ago.
• 15%Named operator — A company we can name stands behind the tool, and we know where it is registered. Anonymous operators are the single strongest predictor of a site you cannot get your money back from.
• 13%Independently listed — How many of our source catalogs carry the tool without us. One directory is a listing; five is a track record no one bought in an afternoon.
• 13%Sourced claims — Every price and ownership claim on the card links the page we read it on. A card with no sources is a card asking to be taken on faith.
• 10%Valid certificate — HTTPS, and the certificate our own handshake came back with: who issued it, and whether it has expired. An expired certificate on a live commercial site says nobody is minding it — so we only say it, and only score it, when we saw that certificate in the past 7 days or a check since it lapsed was refused over it. An older sighting is shown with its date and left unscored.
• 8%Published pricing — What it costs, in a number, before you sign up. "Contact sales" is legitimate; a blank where the price should be is not.
• 7%Visitor reception — What people who signed in to rate it said, smoothed toward the catalog mean until enough ratings land. It steps aside entirely below two.

A check we could not run is redistributed; a check that came back empty is not. Those are different failures and they are scored differently. Ratings nobody has left yet are a young catalog, and that check steps aside. An operator nobody can name is not a gap in our research to be politely excused — it is the finding, it scores zero, and it is the first line under the dial.

Two bands of the scale are ours rather than the reputation checkers’. Their corpus is the whole web, where a 40 is ordinary; ours is a moderated catalog where every listing already cleared the publish gate above, so the distinction worth drawing is between a tool that checks out completely and one where we could not confirm who runs it. A dial reads highly trusted from 85, trusted from 70, mixed signals from 50, and low trust below that.

Where a fact is missing, the wording changes with it. Domain age is read from the registry where we have it; failing that from the launch year we researched, scored a shade lower because it came from the vendor’s own material; failing that from how long the listing has been here, which is a fact about us and is capped, flagged and worded so it cannot be mistaken for the age of the domain. Every line under a dial says which of those it is standing on.

What it is not. It is not a fraud verdict, not a security audit, and not advice. A tool can score in the nineties and still be wrong for you, and a young tool with an anonymous operator can be perfectly honest — it simply has not shown us the things a reader would want to know. One thing a site-reputation service checks and we do not is where the server is hosted; nothing on a card claims otherwise.

Paid placement never moves a tool up an organic list. Not a ranking, not a category page, not a set of alternatives, not a score. Anything commercial is a separate, visibly labelled slot, and if a listing is here because someone paid you will be able to tell without reading a disclosure page.

A listing itself can be bought — see the submission page. The paid tier buys a queue position: a faster review, and no badge requirement. It does not buy a catalog position, and a paid submission clears exactly the same gate above as a free one — including the part where it can be rejected.

Outbound links that can earn us a commission carry rel="nofollow sponsored". No advertiser has ever been shown this arithmetic before it ran, and no vendor can edit their own score, rank or reviews.

We will. Prices change, products pivot, and a fact right in March is wrong by June. A card that is wrong is a bug, and it gets fixed and re-verified rather than argued about — founders and vendors can have their own entry corrected, and so can you.

Report a correctionAbout EffectHubBrowse all categories