Publishing notes
How ChatGPT became our number one traffic source
Dietly's most valuable discovery was not an SEO trick: clear, sourced answers to specific questions are useful to people and easy for AI assistants to cite. That has made assistant referrals a bigger source of attention for us than the generic search traffic we expected at launch.
The referral was a consequence, not a channel hack
Early pages that tried to cover a broad nutrition topic often answered too late. They had the standard pieces: a title, a few reassuring paragraphs, a tool link and a long conclusion. A reader who asked a simple question still had to search inside the page for the answer. That is bad reading experience, and it is bad source material for an assistant trying to support an answer.
The pages that earned citations did something less clever. Their opening paragraph gave a direct answer, included a concrete number where one was appropriate, and immediately named the important qualification. A reader asking whether weight fluctuates could learn that day-to-day changes of one or two kilograms are often water, glycogen, sodium and food volume, then see what to watch over a week.
We made the evidence visible
Nutrition claims are unusually easy to overstate. So the editorial rule became: link a disputable claim to a health body or primary research, and separate what a label says from what a product is supposed to mean. The product examples in a guide use pages from Dietly's catalog, where the underlying data source and fields can be inspected.
That does not make the catalog laboratory truth. Dietly is built primarily on Open Food Facts, alongside community data and clearly labelled estimates. The useful thing is being specific about that limit instead of hiding it behind words such as "verified" or "clean."
Structured pages help machines without writing for machines
We use ordinary semantic headings, compact facts tables, visible review dates and JSON-LD breadcrumbs and FAQs when a page genuinely has questions and answers. Those are not a guarantee of a citation. They do make it easier to identify what a page is about and where a precise answer begins.
The same principle is behind llms.txt. It is a compact guide to the catalog, tools and high-value explainers. It does not replace a normal sitemap or page content. It simply gives systems that look for it a clean map of useful starting points.
The July dip was a useful warning
Referral patterns can change quickly. We saw a July dip, and the honest conclusion was not that a new metadata field would fix it. AI products change their retrieval and citation behavior, and sources compete for attention. Treating a referral stream as an entitlement would be a mistake.
Instead, the response was to improve pages that already earned visits: refresh their sources, correct unclear openings, update their review date when the content actually changed, and check that links still lead somewhere useful. That work improves a page for direct visitors too, which is why it remains worthwhile even if referrals change again.
What we will not do
We will not produce hundreds of near-identical pages to catch query variants, invent author personas or bury the answer under a signup prompt. Those tactics create a catalog that is difficult to maintain and not worth citing. The content plan is deliberately made of individual pages with a distinct question, direction and linking opportunity.
For developers, the same standard applies. A tutorial should show the real request shape, say that /search returns an array, explain rate-limit handling and link the OpenAPI specification. A vague sales page gets neither a useful integration nor a useful citation.