Why AI assistants don’t recommend your brand.

Eight defects, and the fix for each

1. 01

   The page and the structured data disagree

   The page says in stock, the markup says out of stock. The page says €390, the feed says €450. An assistant reading both has to pick, and it may pick the one that loses you the sale — or decide the source is unreliable and quote a competitor instead.

   Generate the structured data from the same source as the visible copy, so a contradiction becomes impossible rather than merely unlikely.

2. 02

   The crawlers are blocked, usually by accident

   A robots.txt that allows everything with a wildcard still gets a brand skipped, because several AI crawlers only act on a rule that names them. Others are blocked outright by a CDN rule nobody remembers writing.

   Name the crawlers one by one, decide per crawler, and check the server logs to see which ones actually arrive. There are around twenty worth naming today, and the list changes.

3. 03

   The page only exists after JavaScript runs

   The content is there in a browser and absent in the fetched HTML. Several AI crawlers do not execute JavaScript, so for them the page is an empty shell.

   Server-render anything that matters. The test is one command: fetch the URL and read what comes back, with no browser involved.

4. 04

   The brand has more than one identity

   A legal name in the footer, a trading name in the title, a third form in the feed, two contact addresses. An assistant cannot tell whether these are one company or three, so it hedges, or names none of them.

   One name, one contact address, one description, repeated identically everywhere — pages, feeds, structured data, directories.

5. 05

   Nothing is addressable

   The work, the services and the answers live inside tabs, modals and accordions with no URL of their own. An assistant cannot cite what it cannot link to, so it cites someone who can be linked to.

   One URL per thing worth citing. If it deserves a recommendation, it deserves an address.

6. 06

   The copy is aspirational where it should be exact

   Finest Italian craftsmanship tells an assistant nothing it can repeat. Material, origin, weight, cut and care are what get quoted back to a buyer, and vagueness is what gets a brand skipped in a comparison.

   Write the facts. Keep the romance for the campaign, not for the fields an assistant reads.

7. 07

   There is no file written for machines

   An assistant assembling an answer about a brand has to infer it from marketing pages, and it infers wrong — stale services, a city the company does not have, a client it never had.

   Publish llms.txt for the short version and llms-full.txt for the long one, and keep both in agreement with the site. State what you are not, as well as what you are.

8. 08

   Nobody is measuring it

   Assistant referrals arrive with thin or missing referrer data, so they land in direct traffic and the channel looks like it does not exist. Meanwhile analytics that only count humans miss every agent visit entirely.

   Measure agent traffic server-side, and measure the outcome as orders by source in your own commerce platform. Impressions are not the unit.

Check this site while you are here

Every item above is applied to the site serving this page. You do not have to take our word for any of it.

• /robots.txt — every AI crawler named one by one, CCBot included
• /llms.txt — the short machine-readable version
• /llms-full.txt — the long reference, so nothing has to be guessed
• /agents — written for assistants, not for people
• /sitemap.xml — every page worth citing, with an address

Where this has already worked

A multi-brand retailer running three boutiques under official brand franchises now receives orders with ChatGPT recorded as the traffic source — a channel that did not exist for that shop before the work. The full account is on the Fashion Italia case study. If you retail under a franchise, the specific version of this problem is on the franchise boutiques page.

Questions we get asked

Is this SEO under a new name?

It overlaps and it is not the same. Classic SEO competes for a position in a list of links. An assistant does not return a list — it returns one answer, assembled from sources it trusts enough to paraphrase. That rewards being unambiguous, exact and machine-readable far more than it rewards keyword coverage, and it punishes contradiction in a way a search ranking never did.

How long before anything shows?

The technical defects are fixed in days and take effect at the next crawl, which can be the same week or a month out depending on the crawler. The content work — exact copy, addressable pages, answers to questions the brand never answered — is the slower half and the one that compounds. Anyone promising a date is guessing.

What does the result actually look like?

Orders arriving with an assistant recorded as the source. On the work we did for a multi-brand retailer operating under official brand franchises, that channel went from not existing to producing attributed orders. That is the unit we report in — not impressions, not rankings.

Can I check whether HelloFrenchy does this to its own site?

Yes, and that is the point of saying it here. Fetch /robots.txt and count the named crawlers. Read /llms.txt and /llms-full.txt. Read /agents, which is written for assistants rather than for people. Fetch any page with curl and confirm the content is in the HTML before any JavaScript runs. Every claim on this page is verifiable on the site serving it.