Is there an open source alternative for tracking AI mentions? Twelve repositories read on 20 October 2026, with the licence and the last commit for each

There are real open source tools for tracking whether AI assistants mention a brand, and on 20 October 2026 we read twelve of them from the GitHub API rather than from anybody's list. Ten carry a recognised open source licence, nine have been pushed to in 2026, and the strongest of them describe themselves as self hosted with bring your own key. The code is genuinely free. What is not free is the model call behind every reading and the person who keeps the thing running, and those two lines are the whole of the cost comparison with a paid tool.

We sell a commercial tracker in this category. Every repository below publishes something we do not publish at all, which is its own source code, and that is stated here because it is true and a buyer should weigh it.

The twelve repositories, as their own API records them

Each row is the repository's own published metadata on 20 October 2026: stars, forks, open issues, the licence GitHub identifies from the licence file, and the date of the last push. No README was summarised and no quality judgement is made about any of them.

repository	stars	licence	last push	language
yaojingang/GEORank	505	Apache-2.0	2026-10-14	Python
ai-search-guru/getcito	493	not identified, reported as Other	2026-09-28	TypeScript
oneglanse/oneglanse	223	MIT	2026-10-04	TypeScript
databluedev/searchmirror	110	AGPL-3.0	2026-10-01	Python
alexpospekhov/searchstack-aeo	108	MIT	2026-09-29	Python
paulmassen/seo-playground	69	no licence file detected	2026-10-19	TypeScript
sarahkb125/llm-brand-tracker	58	Apache-2.0	2025-07-23	TypeScript
anyin-ai/aperture	28	MIT	2026-07-21	TypeScript
hellowalt/aeo-radar	26	MIT	2026-07-01	TypeScript
aws-samples/sample-llm-search-citation-analysis	23	MIT-0	2026-10-14	TypeScript
TheCraigHewitt/namedrop	9	MIT	2026-10-18	TypeScript
nibzard/llm-answer-watcher	9	MIT	2025-11-11	Python

For scale, the GitHub search API returned 287 repositories for the phrase used to find AI visibility monitoring and 25 for the phrase used to find LLM brand mention tracking on the same day. The twelve above are the ones with enough stars, a readable licence record or a recognisable publisher to be worth a buyer's afternoon.

The three fields that decide whether you can actually use one

• The licence, read from the licence file rather than the README. MIT and MIT-0 let a company use and modify the code with almost no condition. Apache-2.0 adds a patent grant and a notice requirement. AGPL-3.0 is the one to take to whoever signs off your contracts, because it reaches code you run as a network service and not only code you ship. One of the twelve is AGPL-3.0 and says so in its own description.
• The licence that is not there. Two of the twelve do not carry a licence a machine can identify: one reports as Other and one has no licence file at all. Published source with no licence grants no permission, so those two are reading material until the author adds one or answers an email. Both are otherwise active, and one of them was pushed to the day before we looked.
• The last push date beside the open issue count. Nine of the twelve were pushed to in 2026 and two were last pushed in 2025, one of them fifteen months before we read it. An engine changes its interface and a tracker stops collecting, so an abandoned collector is the one dependency in this category that fails silently.

What the free version costs, in the two lines that are not zero

Several of these projects use the phrase bring your own key, and that phrase is the entire economics of the self hosted route. The collector is free and every reading is an API call you pay for, so the bill moves with prompts multiplied by engines multiplied by repeats multiplied by how often you re-run the set. Fifteen prompts on three engines repeated three times is 135 answer runs for one pass, and a weekly schedule makes that 540 a month before you have looked at a dashboard.

The second line is the person. Somebody writes the prompt set, keeps the brand matching from counting the wrong thing, notices when one engine's responses stop parsing, and reads the answers. That work exists whether you bought software or cloned a repository, and on the self hosted route it is the only thing standing between the data and nobody.

The genuinely cheapest first reading involves neither. Fifteen buyer questions, typed by hand into the apps your buyers use, three times each, labelled in a spreadsheet, costs nothing in software and gives you a dated baseline you own. Do that before you clone anything, because it tells you what to demand from the code.

Open source collector, open weights model: two different things

Open source in this category means the collector and the scoring are published, not that the model answering the question is. Every project above calls a commercial model or reads a commercial app, and the open weights question is a separate one with its own answer, set out in what is actually downloadable and what open weights means. A downloadable model does not help here anyway, because the thing you are measuring is what the hosted apps tell your buyers.

The one limit that applies to the self hosted route and the paid route alike

Most of these projects reach the engines through APIs, and an API answer and a consumer app answer are two different readings of the same question. The buyer you care about opens the app. Vendors differ on this too and the better ones publish which route they use: one commercial tracker states that it captures ChatGPT and Gemini from the consumer web interfaces and Perplexity through its own answer model, and our own method page names the route for each of the five surfaces we read. Ask the same question of a repository before you trust its numbers, and read the code, because with these twelve you can.

Write the route into your own notes either way. A year of history collected through an API is a year of API history, and it does not become app history later.

Where the repositories beat every commercial vendor, including us

On the single criterion of publishing the thing that produces the number, these twelve projects win outright and no paid product in this category comes close. You can read exactly how a mention is matched, whether a brand name inside a longer word is excluded, how a repeat is counted and what happens when an engine returns nothing. We publish our method in prose at our methodology page and we do not publish our source, so a reader is taking our description on trust in a way that a reader of these repositories is not.

What a paid tool is actually buying you is the maintenance, the engine routes you cannot build yourself, and somebody whose job it is to notice when the collection breaks. That is a reasonable thing to pay for. It is not a reason to skip reading the repositories first, and one of them describes itself in its own words as a free alternative to two of the best known paid products in this category.

Questions about open source AI mention tracking

Is there a genuinely free way to track AI mentions?

Yes, two of them. Ten of the twelve repositories we read on 20 October 2026 carry a recognised open source licence and can be self hosted, and the hand method costs nothing but your afternoon. Both leave you paying for the model calls and your own time, which is where the real cost sits either way.

What is the cheapest AI brand monitoring route overall?

The hand audit, because it costs nothing in software and nothing in API credits: type fifteen buyer questions into the apps three times each and label the answers yourself. The next cheapest is a self hosted collector with your own API key, where the bill is prompts multiplied by engines multiplied by repeats. Compare either against a paid quote by counting answer runs rather than by comparing headline plan prices.

Which licence should a company look for?

MIT, MIT-0 and Apache-2.0 are the straightforward ones for internal use, and nine of the twelve carry one of the three. Take AGPL-3.0 to whoever approves your contracts first, because it has conditions that reach software offered over a network. Treat a repository with no licence file as unavailable until the author publishes one.

Will a self hosted tracker read the same answers as a paid tool?

Only if it reaches the engine the same way. Most of these projects call APIs, several commercial products capture the consumer apps, and the two return different answers to the same question on the same day. Write down which route your data came from, keep it fixed, and compare your own numbers against themselves rather than against somebody else's dashboard.

How do I tell an abandoned project from a maintained one?

Read the last push date next to the open issue count, both of which the API publishes. Nine of the twelve were pushed to in 2026 and the two oldest were last touched in 2025. A collector that has stopped being maintained usually keeps running and quietly stops collecting, which is worse than failing, so check the date before you build a report on it.

Clone one before you take a demo

A buyer who has run a hundred readings through a repository they can read arrives at a vendor call with the right questions and a number of their own to argue from. That is worth more than any feature table, and on 20 October 2026 the code to do it was sitting in public under a licence most companies can accept. Read the licence file, read the last commit date, then read the matching logic, and whichever way you end up buying you will know what you are buying.