Last year I watched an AI engine praise a client in one sentence and cite only their competitors in the next. The client was thrilled to be "mentioned." I was not, because a mention you cannot trace back to your own URL is applause with no door behind it.
That gap is the whole problem with AI search, and almost nobody is measuring it. A brand mention is when the model says your name. A citation is when it attributes the answer to a page on a domain you own. Mentions feel nice and disappear. Citations compound, because every one sends the engine and the reader back to a property you control.
We build AI-marketing agents at Humanswith.ai, and we run the same loop for clients every week: measure, produce, optimize, design. This year I open-sourced the core of it. Four small tools, MIT-licensed, no account, no API keys, that are the honest skeletons of the agents inside our hosted Workspace. The Workspace runs the loop at scale. The free tools hand you the method.
**TL;DR** — one tool per step:
`ai-visibility-probe-lite`
— measure mention-share versus citation-share across engines.contentos-agent-lite
— write from an eight-gate process with canonical-first distribution.aeo-site-audit-lite
— audit the three retrieval gates: fetchable, chosen, extractable.brand-card-lite
— generate on-brand social cards from a tokens file.Here is the loop, one tool at a time.
You find them by running brand-free discovery prompts through the engines and recording who gets named and who gets cited. The first measurement is almost embarrassingly simple: ask the questions your buyers actually type, then read the answers.
The first tool, ai-visibility-probe-lite
, is a kit for exactly that. You write brand-free discovery prompts, the questions a buyer asks before they have heard of you ("best X for Y," never "is Acme any good?"), run them yourself in ChatGPT, Perplexity, or Gemini, and paste the answers back. It then reports two numbers that most dashboards collapse into one: mention-share, how often the answer named you, and citation-share, how often it cited a URL you own. Brand-free prompts are the discipline here. Asking "is my brand good" measures how the model feels about you, not whether it treats you as a source. [1]
The first time you run this on your own company, it stings a little. That sting is the point.
You produce citable content by writing from a documented process with sources, not from a single prompt. The fastest way to write a page no engine will ever cite is the opposite: paste a topic into a chatbot and ship the first draft. The second tool exists to stop that.
contentos-agent-lite
is a content agent you run inside your own coding assistant. It writes from a documented process rather than a single prompt, walking eight gates: business context, research, a source pack, a brief, a draft, an editorial pass, a publish-readiness check, and distribution. It will not invent a fact that has no source, which already puts it ahead of most AI writing.
Its eighth gate is the one teams skip and then pay for. Gate 08 is canonical-first distribution. Before anything ships, it lints the page for the signals that decide whether an engine can fetch and attribute a citation: one canonical URL, one H1, Open Graph tags, Article structured data. Then it drafts the LinkedIn, Medium, and dev.to versions, and every one of them links back to your canonical page. Publish on your own domain first. Make every rented copy point home, so the authority piles up where you own it instead of where you borrow it. [2]
A page can be beautifully written and still never get cited, because three gates sit in front of every citation, and most people only ever check the first.
The third tool, aeo-site-audit-lite
, checks all three. Fetchable: can a crawler reach and index the page (HTTP status, robots and noindex, canonical, structured data). Chosen: among the pages it can reach, why does it pick a competitor (you hand it the URLs an engine cites today, and it classifies the gap as authority, freshness, a competitor's data table, or missing proof). Extractable: once chosen, can it lift a clean answer out (a concise lead answer, headings, schema, lists and tables it can quote in one block).
Point it at a URL or a local file and you get a scored report and a prioritized fix list. It runs offline, needs no keys, and uses only the Python standard library, so you can read every line before you trust it. [3]
You keep it on brand by generating every social card from one brand-tokens file, so colors, fonts, and logo stay consistent wherever a post lands. Citable content still has to look like you when it shows up in a feed. The fourth tool, brand-card-lite
, turns a tiny brand-tokens file (your colors, fonts, and logo) into a self-contained, on-brand social card, and it lints those tokens for contrast and consistency so the text actually meets WCAG and the fonts have real fallbacks. No image model, no cloud service, no keys. Templating and a little math you can audit. [4]
Across the audits we run, the same four mistakes show up again and again, and each tool above targets exactly one of them.
| Loop step | Free tool | The mistake it fixes |
|---|---|---|
| Measure | `ai-visibility-probe-lite` |
Stopping at "are we mentioned?" instead of measuring the citations to URLs you own. |
| Produce + Publish | contentos-agent-lite |
Shipping a one-prompt draft with no sources or structure — the page an engine skips. |
| Optimize | aeo-site-audit-lite |
Never checking retrieval, so a page stays unreachable, un-chosen, or un-extractable. |
| Design | brand-card-lite |
Letting the silo win — distributing with no canonical home, so a rented platform takes the authority. |
You do not need the hosted product to start. Here is the minimum loop, the one I would run if I had an hour:
People ask why I open-sourced the part that took us years to figure out. The honest answer is that the method is not the moat. Brand-free measurement, the mention-versus-citation split, the eight content gates, the three retrieval gates, canonical-first distribution: that is the part that changes outcomes, and it is now fully in the open.
The hosted Humanswith.ai Workspace adds the part you cannot do by hand at any real scale: automated multi-engine scans on a weekly cadence, the publishing and re-measurement loop that proves whether a fix actually worked, and the team and hosting around it. The free tools tell you where you stand. The Workspace runs the loop for you. [5]
It is a deliberate bet. The companies that win AI search will treat it as an operating loop, not a one-time audit. So clone a tool, measure one prompt set, and fix one page this week. That is the entire ask.
Are these really free and open-source? Yes. All four are MIT-licensed on GitHub, with zero runtime dependencies and no API keys. You can read every line before you run it.
What is mention-share versus citation-share? Mention-share is how often an answer names your brand. Citation-share is how often it cites a URL on a domain you own. An engine can name you while sourcing only competitors, so the tools report both. [1]
What does canonical-first distribution mean? Publish on your own domain first as the canonical version, then adapt for the platforms, with every copy linking back to that URL. The citation authority stays on the property you control instead of the silo you rent. [2]
Will this guarantee AI cites me? No, and anyone who promises that is selling something. The tools measure the gates that come before a citation and give you a way to close them in priority order. The result is earned.
When should I move to the hosted Workspace? When you want the loop run for you across a real content program, rather than running each tool by hand. [5]
[1] `[ai-visibility-probe-lite](https://github.com/humanswith-ai/ai-visibility-probe-lite)`
(MIT)
[2] `[contentos-agent-lite](https://github.com/humanswith-ai/contentos-agent-lite)`
(MIT, incl. gate 08 canonical-first distribution)
[3] `[aeo-site-audit-lite](https://github.com/humanswith-ai/aeo-site-audit-lite)`
(MIT)
[4] `[brand-card-lite](https://github.com/humanswith-ai/brand-card-lite)`
(MIT)
[5] [Humanswith.ai Workspace](https://humanswith.ai/platform/)
Originally published on gregshevchenko.com.