The Legibility Sprint
A seven-day protocol for making a person or company readable to AI systems. Published in full, free to run.
Search used to index pages. AI systems reconstruct you from fragments, and if the fragments are thin, stale or tangled with someone else's, the answer gets built without you. Not because you are worse than the person named instead. Because you are harder to read.
This is the protocol I ran on myself before I ran it for anyone else. It was a paid PDF until August 2026. It is now published in full, for a plain reason: a method that models cannot read is a poor advertisement for legibility work.
Two definitions before the steps, because everything below depends on them. Visibility is how often you are mentioned. Legibility is whether a model can identify you correctly, place you in relation to the things around you, and attach a claim to you without taking on risk. Most people who feel invisible have a legibility problem, not a volume problem.
Day 1 Measure what the models say now
- Write down a fixed list of questions and do not change it later. Three kinds: identity (who is [name]), category (who works on [your field]), and recommendation (who should I hire for [your problem]). Eight to fifteen questions is enough.
- Run every question in ChatGPT, Claude and Perplexity. Run each one three times, in a fresh chat, with no personalization and no memory. Answers vary between runs, so a single run tells you almost nothing.
- Save the raw answers, not your summary of them. Note the date and the model version.
- Score each answer on four things: does it recognize you at all, is the description accurate, does it attribute your work to you, and would it recommend you. Count how many of the runs got each one right.
- This is your baseline. Everything later is measured against it, and you never rewrite it.
Caveat. Nothing here is optional. If you skip the baseline you will have opinions later instead of results.
Day 2 Write your llms.txt
- Create a plain text file at yourdomain.com/llms.txt. It is a summary of who you are, written in the order a machine reads rather than the order a human browses.
- Open with one sentence that defines you: name, role, what you do, for whom. This exact sentence goes everywhere else too, word for word.
- Then list, in short blocks: your work and what each thing is, any terms you coined with a definition and a date, your other profiles, and how to contact you.
- Write every claim so it can be checked. "Live pilots in two markets" is checkable. "Industry leader" is not.
- Mine is public at katyashalel.com/llms.txt. Copy the structure, not the content.
Caveat. llms.txt is an emerging convention, not a standard anyone is obliged to honor. It costs an hour and it is read by several crawlers today; treat it as cheap insurance, not as a guarantee.
Day 3 Deploy structured data
- Add JSON-LD to your key pages. These are invisible labels in the page code: a human sees a paragraph, a machine sees "this is a person, this is her role, this is her company."
- Minimum set: Person on your about page, Organization on your company page, Article or BlogPosting on every essay, and DefinedTerm anywhere you define a term you invented.
- In the Person block, fill sameAs with every profile that is really yours. That is the property that tells a model the scattered mentions are one entity.
- Keep the same one-sentence definition in the description field. Same words as llms.txt. Consistency is the signal; variation reads as uncertainty.
- Validate the output before you ship it. A broken JSON-LD block is ignored silently, and you will never know.
Day 4 Put your entity in the public record
- Models look for confirmation outside your own website. A claim only on your site is a claim; the same claim in an independent record is a fact they can use.
- Create or correct your Wikidata item. It needs a label, a description, an occupation, and sourced statements. Notability rules are real, so cite external coverage rather than your own pages.
- Bridge your names. If you appear as two spellings or in two alphabets, connect them explicitly in Wikidata, llms.txt and your sameAs list, or they stay two different people.
- Make sure your professional profiles carry the same defining sentence. A profile that describes an old role will keep winning retrieval long after the role ended.
Caveat. If Wikidata declines the item, do not force it. Use the alternative track: consistent profiles, third-party mentions, and structured data on your own domain.
Day 5 Build the page a machine can read
- Make one page that answers, in plain sentences, the questions from your Day 1 list. Not marketing copy. Answers.
- Structure it so any single paragraph makes sense pulled out on its own, because that is how it gets used.
- Define your terms on the page, each with a definition, an example and a date.
- Check robots.txt actually allows AI crawlers. Many templates block them by default and nobody notices.
Day 6 Put it where it gets picked up
- Structured content gets indexed fastest where indexing already happens: your own domain first, then the platforms that publish quickly and get crawled often.
- Repeat the same defining sentence on every platform. Not a variation. The same words.
- If you coined a term, define it once in a dated public place and link every later use back to that definition, so the term stays attached to you.
- Get at least one independent source to describe you the way your own pages do. One corroborating source outweighs ten of your own pages.
Day 21 Re-measure
- Wait. Retrieval and indexing move on their own schedule and a week is not enough to see anything.
- Re-run the exact Day 1 questions, same three models, same three runs each, same scoring.
- Report the delta as a share of runs, not as a screenshot. "Named in 7 of 9 identity runs, up from 2 of 9" is a result. One good screenshot is an anecdote.
- Log everything that changed in the period, including things you did not control, so you do not credit the protocol for a model update.
- Then repeat quarterly. Models retrain, competitors publish, sources decay. This does not stay done.
What you have at the end
- A live llms.txt on your domain.
- JSON-LD deployed on your key pages: person, organization, articles, defined terms.
- An entity record in the public graph, with your name variants bridged.
- One page written to be retrieved rather than browsed.
- A frozen question set and two dated measurements against it, twenty-one days apart.
- A repeatable quarterly routine, because none of this stays done.
What this protocol cannot do
- No one can guarantee a position in an AI answer. Anyone who does is selling astrology.
- Model answers are not deterministic. The same question can return different answers in the same hour, which is exactly why the protocol asks for repeated runs and a fixed question set.
- Results are specific to the markets, languages and questions you tested. A gain in Russian-language queries says nothing about English-language ones, and vice versa.
- Structure amplifies substance, it does not replace it. If there is nothing verifiable to find about you, legibility work makes the absence easier to read.
Questions
Why is this free?
Because a method that models cannot read is a bad advertisement for legibility work. Published in full, the protocol is citable, checkable and repeatable by anyone, which is worth more to me than the price of a PDF.
Do I need to write code?
No. Every step is text, files and settings. An AI assistant can execute all of it; the part that does not delegate is deciding which entity is canonical and which claims you are willing to stand behind.
How long does it really take?
The work is roughly seven working sessions, but the calendar is longer: the day 21 re-measurement is part of the protocol, and quarterly repeats are part of keeping the result.
What if I have several companies or brands?
Then the ordering matters more than any single step. You have to decide which entity carries the authority and which ones inherit it, and get that wrong and you spend a quarter reinforcing the weakest one. That is the case where people usually ask me to run it with them.
Is llms.txt an official standard?
No. It is an emerging convention that several crawlers read today and no one is obliged to honor. It is cheap to add and easy to remove, which is why it is in the protocol despite the uncertainty.
Want this done for you
The protocol above is the whole method, free, and plenty of people run it themselves. If you would rather not, or your setup has several connected entities and the ordering matters, tell me what you are working with. I read every enquiry myself and reply within one business day.