Start here

Below is the order I did this in, and under each step what it costs to do that step at the wrong moment. Each of those costs is one I paid. If you would rather tick boxes than read reasons, take the checklist.

One thing comes before step one. You have to be able to say what you are building and how it differs from what is already on sale. If you cannot say it yet, start there: where a niche starts when everything is taken.

Everything below assumes you have that answer. And this is my log rather than a method: your numbers will come out different, but the order is the part I hope saves you a month.

Hand it to an agent

/atlas:start — walks a project through the route, runs the audits at the right points and names three things to do this week

The command is not in Claude Code out of the box — how to install the plugin.

What we are solving

The project has been live for weeks. The only visitor in the logs is you.

You have read a lot of advice. It gave you twelve things to do and no idea which one is first, and I picked the wrong one and paid for it in weeks.

Order is the whole problem. Each item below is cheap on its own. Do them in the wrong order and you lose the month, because the work lands before the thing it depends on and pays nothing back.

The first item is not technical. You find out whether anybody searches for this, and then whether the people who search pay for anything at all.

The gates on this route stand before the work, not after it. If the project intends to charge, some of them move right to the front: break-even named as a whole number of people costs an evening to work out, and forty written pages to find out the hard way.

Steps

The section numbers are a reading order. The doing order crosses them twice, and I marked each crossing where it falls.

First — does anybody search for this, and do they pay for anything

1. Collect the exact wordings people use, from support, reviews and forums — the words people actually type.
2. Check the Russian wordings in Yandex Wordstat, with the operators applied, because without operators you get the category total rather than demand for your wording. No Russian-speaking audience? Skip to step three — same page.
3. Check the English wordings in Google. Read the suggest dropdown in both engines — same page.
4. Search the platform your product lives in, from a fresh account — same page.
5. Read a zero on your most obvious phrase as a finding, not a glitch. Either nobody has this problem, or you invented the word — same page.
6. Name the paid competitor on your main query. Write down the shape of their price: per seat, per action, per month — whether that audience pays.
7. Count the ads sitting over the query yourself. Write down the free path a person has today — same page.

This takes a day or two, all of it in a browser, and it decides whether the rest of the route is worth doing at all. Everything below assumes somebody types these words and pays for something.

Once there is a live domain — can a crawler read the site, and does anyone know it is there

8. Run curl -sL on your own page and look for the body text in the response, because text that is missing there is text the crawler never sees — Google does not see your site.
9. Read robots.txt on the live domain, every line of it, because one line closes the whole domain: a Disallow: / that came over from a staging config — same page.
10. Check noindex in the meta tag and in the X-Robots-Tag header. It drops the page from the index on purpose, and no browser shows you that header — same page.
11. Check that the canonical on a few pages points at those pages. It names the real URL, so aiming it at the home page asks search to discard the rest — same page.
12. Request one page from outside your network, without cookies. Your own network is trusted, so an edge rule that blocks the crawler stays invisible from home — same page.
13. Verify a domain property in Search Console and submit the sitemap once. A URL-prefix property covers only the form you typed, and its report then stays empty — submit and verify.
14. Import the property into Bing Webmaster Tools. Its index also feeds Copilot answers, and those readers are a separate audience — submit and verify.
15. Serve User-agent: * and Allow: /, and name an agent only to block it: an allow-list expires the first time a vendor renames something. This site has never named one to allow it — AI crawlers and llms.txt.
16. First crossing. Tag every outbound link and store the source on first contact. A messenger or a store passes no source unless you ask for it — where the user came from.
17. Name the single action that counts as activation. Activation is the one thing a user does that proves the product worked — the numbers worth reading weekly.

Step sixteen belongs to section six, and you have to do it now. You cannot reconstruct attribution afterwards.

Every untagged day is a day of arrivals nobody will ever classify, however good the dashboard gets later. My own launch day is one of them.

Once the crawler is let in — will the page be found, and will it be quoted

18. Group the wordings you collected into clusters, one URL each. A cluster is a set of wordings that want the same page — what to write.
19. Order the queue by money: transactional, then comparison, then informational. That is ready to buy, then still choosing, then still reading. I ran out of energy long before I ran out of clusters — what to write.
20. Write the first page in the four-block shape — what a page is made of. The blocks: the problem, the steps, what did not work, how to check.
21. Put the answer in the first three sentences of that page. That block is the passage a model lifts — what a page is made of.
22. Add JSON-LD with sameAs listing your profiles and listings. JSON-LD is a block of machine-readable facts about the page. Google documents it for the Knowledge Graph, and for AI citation nobody has measured it — why AI answers cite someone else.
23. Publish llms.txt only if you ship developer docs, and name only pages that are actually written. My own log records 0 fetches by any AI agent in 16 days — AI crawlers and llms.txt.
24. If the product lives in a platform, put its function in the name. Platform search matches the name, not the description — search inside the platform.
25. Take the first weekly snapshot and append it to a file — the numbers worth reading weekly.

Once there is something worth linking to — who links to you, and what does a user cost

26. Write the listing card once: name, description, category, screenshots, link — where the first links come from.
27. Check each venue is alive before you spend an evening on the batch, because published lists of directories go stale. Then submit in small batches, with tagged links — where the first links come from.
28. Write to one listicle author whose page already ranks above you — where the first links come from.
29. Second crossing. Re-run the citability work now that something links to you — why AI answers cite someone else.
30. Record the cost of every action on its own event row. I never had such a column at all, so every report I built reconstructed the cost from a rates table afterwards — what one user costs.
31. State break-even as a whole number of payers per month — what one user costs.
32. Read the series and fix the earliest large drop, not the last one — the numbers worth reading weekly.
33. Post where the shortlists get written: answer the question and say the product is yours. Never ask for upvotes — communities and forums.
34. Put who, what, when, source, cost, client and outcome on every event row. Cost and source can only be written at insert time, so it really belongs next to step sixteen — what an event row has to carry.
35. Name a price. Put the wall after the first real result, and ask one person for money. On my dashboard, nobody paying looked exactly like nobody being asked — getting the first person to pay.

Step twenty-nine goes back to section three. On-page citability pays once a model has seen your name somewhere else, and by this point that starts being true.

What did not work

These are mistakes of order. I did each of them correctly, at a moment when the work could not pay, so the fix was never to do the thing better but to do it later.

• Optimising for a phrase I invented. My word for the core feature was not the word people typed. Pages, links and dashboards all worked, and the demand walked straight past them.
• Backlinks while the site was still blocked. I spent a week on directory submissions, and the pages behind those links carried a noindex header the whole time. Search dropped them about as fast as I filed them.
• Writing before the cluster map existed. Two of my pages chased the same intent, search alternated between them, and both stayed flat. I merged them, added a redirect, and paid for that text twice.
• Tagging links after the launch. The launch day arrivals came in untagged and stayed that way. No later analysis brings them back, because there is nothing left to analyse.
• Citability work before crawler access. The pages scored well and not one agent had fetched them. The edge refused every request while I polished the opening sentences.
• Cutting costs before there were payers. I carefully optimised a few dollars of fixed cost, with zero payers on the other side. When nobody pays, there is no break-even to move.
• Publishing a step I had already measured as empty. Step twenty-three said publish llms.txt. The log two clicks away recorded not one AI agent fetching it, and the measurement and the instruction lived on the same site for a week.
• Telling readers to allow agents by name. Step fifteen did exactly that, while this site’s own robots.txt was a wildcard the whole time. My own file was the better advice, and I never read it as advice.
• Following the section numbers literally. They are a reading order, not a schedule. Read them as a schedule and attribution lands after distribution. That is one section too late.

Verify

Each stage ends with one sentence you should be able to say out loud. If you cannot say it, stay in the stage, because the next stage is built straight on top of it.

• Demand. A file of wordings, each with its origin and the engine it was checked in. A named paid competitor, or a stated reason the market looks empty.
• Access. URL Inspection says the URL is on Google. A link you published today carries a source tag, and that tag lands in the database.
• Pages. Every cluster has exactly one owning URL. The first page answers its own title in its opening sentences. The snapshot file has a row in it.
• Links and money. At least one listing whose HTML contains your link. One recorded cost on a real event row. A break-even you can say as a whole number.

When something specific is broken, start from the symptom instead of the stage: Tools.