How we review AI tools

This page explains what we list, where entries come from, how we build and check profiles, how the order of results is set and how we check reviews. The rules are the same for every entry. No vendor can change them with money.

What we list

ai-toolfinder lists AI tools for business in 50 categories. Every entry has exactly one type and one main category, which describes what it is used for. The types are:

Tool: software you buy or use as a service.
Repo: an open-source project whose code is public, for example on GitHub.
Model: an open or commercial AI model.
Dataset: data for training or evaluating AI.
API: an AI interface for developers.
MCP server: a server that connects AI assistants to apps and data through the Model Context Protocol.
Skill: a skill, GPT or plugin for an AI assistant.
Extension: an AI extension for the browser.
Paper: a research paper with published code.

An MCP server for a CRM, for example, has the type MCP server and the category Sales. We list an entry if it meets these conditions:

It serves a business purpose that fits one of the 50 categories.
Its AI function is described on a public page, usually the vendor's or the project's own page. We link to that page as a source.
The page is reachable and the information can be checked there.
It can be used today. Announced or beta features do not count.

Anyone can suggest an entry. Suggestions are subject to the same conditions. Where a profile contains an assessment from the perspective of the team using the tool, it is marked as opinion and does not count towards the stars.

Where entries come from

We find candidates in open sources: official interfaces such as the GitHub API, the Hugging Face Hub, the official MCP registry and arXiv, announcements and feeds from vendors, submissions through our form, and our own observation of new and trending tools. Other directories serve at most as a hint that a name exists. We never take descriptions, ratings or other content from them. Every fact comes from the vendor's or project's own pages or from another primary source. We never bypass bot protection such as captchas or access checks. If a page can only be reached that way, we do not use it.

How profiles are built and checked

We build profiles with the help of AI. A language model drafts the profile from the pages we fetched, and every core fact (for example pricing, data hosting, certifications, interface languages, headquarters and AI features) needs its own source. Automatic checks then compare each fact with its source, look for duplicates of existing entries and test the text against our quality rules. What fails a check is corrected or left out. People check random samples of every batch. Nick is responsible for the content. The date "Profile checked on" shows when we last compared the profile with its sources.

New entries go live in daily batches. We publish at most 30 new pages per day that search engines may index, so that every batch can be checked. Comparison pages, alternatives pages and guides do not count towards this limit. When we open new categories, we may publish one larger starting batch so that no category looks empty.

When a page is open to search engines

Every entry has its own page. Search engines may index it only when the entry has enough substance:

Tools, APIs and browser extensions: a full profile with a description of at least 120 words, at least two AI features, at least one source and a source for every core fact that is filled in.
Open-source projects: at least 500 stars on GitHub and activity in the last 90 days.
Models: a known provider or at least 1,000 downloads per month.
Datasets: at least 1,000 downloads per month and a stated license.
MCP servers: maintained by the vendor of the connected app, or at least 100 stars.
Skills and plugins: a known provider or at least 100 stars.
Research papers: a linked code repository that itself meets the threshold for open-source projects.

We check these values after the first weeks and adjust them if needed. Entries below a threshold can still appear in lists, but their own page is closed to search engines.

Besides full profiles there are short entries. They meet the same conditions and show the headquarters, category, a short description and one AI feature, each with a source. Pricing, data hosting and certifications are still missing. Short entries appear in the lists, but their own page is closed to search engines. We mark short entries as "not yet reviewed" when we could not yet confirm the company's age or size. Because they have fewer details, short entries usually rank below full profiles in the "Relevance" order (parameter completeness, see below). Vendors can claim a short entry and complete it.

Keeping entries up to date

We check entries again at regular intervals, more often where things change quickly, such as pricing pages or active open-source projects. The checks look for changed prices, broken links, moved pages and projects without activity. If a product is discontinued, its website no longer works or a project has been abandoned, we archive the entry. Its page then redirects to a successor or to its subcategory. The date on each profile shows the last check.

Vendors can claim their profile, see For vendors. After that they can change the details of their entry, each with a source. We check changes before they go live, and facts changed by the vendor are marked as confirmed by the vendor. Anyone can report an error, see "Reporting errors" below.

What "Strong in a country" means

"Strong in" names the markets where a tool is strongest, not every country where it is available. Many tools, Workday or ADP for example, are sold in far more countries. We only list a country if we find evidence of particular strength there, for example:

the vendor's headquarters or home market,
a documented share of customers or revenue in the country,
market leadership according to a market study,
own data centers or official certifications in the country,
a dense network of locations,
many named reference customers in the country.

An office, a country page, the local language or local payroll alone do not count. We list at most eight countries per tool, the strongest first, and state the evidence on the profile. "Available internationally" means the tool is also offered in other countries. In countries with their own language, this includes tools whose interface offers that language, for example Spanish for Spain, Mexico and Argentina, or German for Germany, Austria and Switzerland. The language does not count as strength.

How the order is set

We publish how the order is set so that anyone can check it. German law (Section 5b(2) of the Act against Unfair Competition, UWG) also requires us to disclose the main ranking parameters and their weighting.

The default sort order is "Relevance". It uses four parameters in a fixed order of priority:

Country fit, only if you have selected a country: first tools where the country is among the first three countries with evidence of strength, then tools where the country is listed as strong further down, then tools that are available internationally or whose interface offers the local language. If you select two or three countries, these levels are added up across the selected countries.
Bayesian average of the reviews (explained below).
Profile completeness: how many of these ten details are filled in: a description of at least 120 words, at least two AI features, country fit with a reason, headquarters, data hosting, pricing model, company sizes, interface languages, certifications, at least two sources.
Name, alphabetically.

On weighting: a later parameter only decides if all earlier ones are equal. Country fit therefore comes before everything else. The Bayesian average decides within the same country-fit level, completeness only if the Bayesian average is equal, and the name only if everything else is equal.

The sort order "Rating" uses the Bayesian average, and the name if two tools are equal. "A to Z" sorts alphabetically by name. "Most countries" sorts by the number of countries with evidence of strength, and the name if two tools are equal. Reviews and payments play no role in these two sort orders.

The Bayesian average stops a tool with a single five-star review from ranking above a tool with many good reviews. The formula is: (C × m + sum of all stars of the tool) divided by (C + number of its reviews). C = 3, and m is the average of all published reviews on ai-toolfinder.com. As long as no more than 30 reviews have been published in total, m is fixed at 3.8. Examples with m = 3.8: a single five-star review gives 4.1. Thirty reviews averaging 4.5 stars give 4.44. A tool without reviews sits at 3.8. All published reviews count equally, whatever their verification level. In practice, a tool without reviews has a Bayesian average of exactly m. While m is fixed at 3.8, all tools without reviews are equal on this parameter, so the order among them is set by profile completeness and then by name.

The star average shown on a profile is the simple average. Because the order uses the Bayesian average, it can differ from the average shown.

Filters decide which tools are shown. A search term does two things: it shows only tools that contain all of its words, and in the sort order "Relevance" it comes first, before the four parameters above. A word also counts as contained in its singular form and when a related term from our list of synonyms appears, for example "admin" also finds "HR administration" and "HRIS", and "payroll" also finds "Lohnabrechnung" and "nómina". If a word of five letters or more finds no tool at all, we also accept words that differ from it by one letter, or by two letters for words of eight letters or more. Search words of up to three letters only count at the start of a word. The levels are: exact tool name, name beginning with the term, all words in the name, all words in the name or parent company, one word in the name, all words in the short description, all words in the categories and their keywords, then all other matches (AI features, tasks, AI use cases, certifications, EU hosting). Within a level, tools found only through a related term come after direct matches, and tools found only through a typo come last; related terms do not count for the levels based on the name. Tools that match all other filters but are neither strong nor available in the selected countries are hidden, and we show how many there are; on request we list them in a separate section. Payments, click counts, a claimed profile or invitations by vendors have no influence.

Questions of three or more words on the home page are translated by the language model Claude (Anthropic) into the same filters you can set yourself (task, country, company size, price, free plan or trial, EU data storage); after that exactly the rules of this section apply. Claude does not pick tools and does not change the order.

"Alternatives to …" pages list tools in the same niche. The tools closest in substance come first. We compute closeness once per data update with a Google language model (Gemini embedding) from description, AI features and profile, rounded in steps of 0.05. With equal closeness, shared strong countries decide, then profile completeness.

How we check reviews

As required by Section 5b(3) UWG, we explain whether and how we make sure that reviews come from people who have used the tool.

Nick reads every review before it is published. He checks whether it meets the review guidelines, whether own use is evident and whether there are signs of a conflict of interest. We do not check whether every single statement is correct. Reviews reflect the opinions of their authors.

Every review shows its verification level:

Email verified: the person signed in with a link we sent to their email address. This applies to every review. It shows that the person has access to this address, not that they have used the tool.
Work email verified: the email address belongs to a company domain and not to a free email provider. This suggests a professional background but does not prove use.
Use verified: Nick has seen a screenshot from the tool while signed in to it, or an invoice, and accepted it as proof. We delete the proof when we decide, at the latest 30 days after the decision. This level exists if the function is offered.

Only the level "Use verified" shows that the person has used the tool. For the other levels, we check plausibility and conflicts of interest but cannot guarantee use.

Automatic checks support moderation:

The form rejects reviews from an email address at the domain of the reviewed tool, and a second review of the same tool from the same account.
The system flags addresses of competitors in the same niche and unusually many reviews in a short time for a closer look.

Reviewers confirm that the review is based on their own experience and that they have no connection to the vendor or a competitor. If a vendor invited the reviewer or there was an incentive, the review says so.

We publish negative reviews under the same rules as positive ones. Vendors cannot have a review removed just because they do not like it. We do not write reviews ourselves and do not buy any. To get the first reviews, Nick asks people from his professional network who use a tool for their assessment, without anything in return and without any requirements on the content.

Comparison pages and product images

On comparison pages we set two tools side by side using the same facts shown in their profiles, each with a source. We only compare pairs that people actually search for.

Where a comparison page shows images, they show the product itself, not the vendor's website. We take them from pages the vendor has made public, such as product pages, help articles or feature announcements, and name the source and the date of capture under each image. The images serve as evidence for the comparison (quotation under Section 51 of the German Copyright Act). We only show images when we find one for both tools, so neither tool looks better just because it has a picture. Vendors who would like a more recent image or object to an image can write to us through the contact form.

Independence

ai-toolfinder.com is operated by Nick's Advisory. Nick advises companies on AI projects in HR. The directory is kept separate from this:

No vendor pays for being listed, for texts, reviews or the order of results.
At launch there are no paid placements and no affiliate links. The redirect to vendors only counts clicks.
If Nick's Advisory works for a listed vendor for pay, this is stated on that vendor's profile.
Should there ever be paid placements or affiliate links, we will label them visibly. Paid placements carry the label "Ad".

Reporting errors

If something in a profile is wrong, use "Report an error" on the profile or write to nick@nicksadvisory.com. Name the tool, the incorrect information and, if possible, a source. We check it, correct the profile and update the check date. Vendors can also claim their profile and suggest corrections directly, see For vendors. Please report illegal content via Report content.

Last updated: September 2026