Serving AI crawlers a simplified markdown copy of your pages does not appear to buy you citations. Digiday's Media Briefing of 6 August 2026 sets out the case for and against the practice and reports that GEO platform Promptwatch examined over 1.6 million citations in AI answers, finding that markdown did not increase the likelihood of content being cited.
The case for markdown is about cost, not visibility
TollBit co-founder and CEO Toshit Panigrahi is cited in the Digiday briefing of 6 August 2026 saying the company sees an average 90% reduction in tokens when content is converted to markdown, which cuts the compute cost of crawling and inference. That is a genuine number about a genuine problem. It belongs to whoever is paying for the compute, and it says nothing about whether an AI assistant will quote you.
The crawl load behind the argument is not hypothetical either. Digiday's 6 August 2026 briefing cites H1 2026 figures showing roughly 4x higher median AI scraping on European sites than on North American ones.
The evidence against a citation lift
Promptwatch supplied the counterweight: more than 1.6 million citations in AI answers examined, with no increase in citation likelihood for markdown. That is the largest number anyone has put on the table so far, and it points the opposite way to most of the advice published in 2026.
Read it for exactly what it is. Promptwatch looked at citations that had already happened, not at the same pages served both ways under controlled conditions. It is a single third-party study, so it is strong evidence that the effect is not obvious, not proof that markdown can never matter.
Oasy, which tested ad placement in both formats, reported through co-founder Choy Travers that ads in markdown had not changed advertisers' results or win rate yet. Publishers quoted on the sceptical side of the briefing raise a separate objection: markdown removes the friction that previously protected content from scraping.
What Google says about markdown and llms.txt
Digiday's 6 August 2026 briefing notes Google's guidance that markdown is not needed. Google has also clarified that maintaining LLMS.txt or similar files will neither harm nor help visibility or rankings, because Google Search ignores them. So the file sitting on your server is not a risk. It is also not doing the job some 2026 checklists claim it does.
Who tested what
The four positions in the briefing measure different things, which is why they can all be true at once. Here is what each party reported and what kind of claim it is.
| Who | What they reported | Type of claim |
|---|---|---|
| TollBit (Toshit Panigrahi) | Average 90% reduction in tokens when content is converted to markdown | Crawl and inference cost |
| Promptwatch | Over 1.6 million citations in AI answers examined; markdown did not increase the likelihood of being cited | Citation study |
| Oasy (Choy Travers) | Ads in markdown had not changed advertisers' results or win rate yet | Ad performance |
| Markdown is not needed; LLMS.txt and similar files neither harm nor help visibility or rankings, because Search ignores them | Platform guidance |
What to do with a markdown project on the roadmap
Much of 2026's GEO advice has assumed markdown mirrors and llms.txt files improve the odds of being cited by AI assistants. The Promptwatch dataset is the first sizeable body of evidence pointing the other way, so treat the citation benefit as unproven until someone runs the controlled test. Developer time spent building a mirror is time not spent on crawlability, clean page structure and claims an answer engine can lift cleanly out of the page.
That ordering holds whether you are working on GEO or on broader AI SEO. Pages that already exist and already rank are the cheapest thing to fix. Making the facts inside them extractable, sourced and unambiguous is content work, not a build project, and nothing in the briefing suggests a format conversion substitutes for it.
What this means for Thai marketers
Thai publishers and brands tend to run smaller development teams than the US and European publishers quoted in the briefing, so a markdown mirror is a heavier commitment here relative to the payoff on offer. There is also nothing in the reported data about Thailand. The 4x scraping comparison covers European versus North American sites in H1 2026, so it says nothing about how hard AI crawlers hit Thai domains. The reasonable position is to treat the mirror as a cost decision about crawler load rather than a visibility tactic, and to spend the same hours on structure and sourcing inside the pages you already publish.
Frequently asked questions
Should I delete my llms.txt file?
There is no need to delete it. Google has clarified that maintaining LLMS.txt or similar files will neither harm nor help visibility or rankings, because Google Search ignores them. Deleting it changes nothing for Google Search, and the briefing does not state how other AI systems treat the file.
Does serving markdown to AI bots get me cited more often?
The available evidence says no. Promptwatch examined over 1.6 million citations in AI answers and found that markdown did not increase the likelihood of content being cited. That is one third-party study of citations rather than a controlled experiment, so treat it as the best evidence currently published, not a settled result.
Is markdown for AI crawlers useless then?
No, it has a documented cost benefit. TollBit reports an average 90% reduction in tokens when content is converted to markdown, which lowers the compute cost of crawling and inference. The open question is who that saving belongs to, since it is a cost claim rather than a visibility one.
Does markdown help or hurt ad revenue in AI surfaces?
Oasy reported no change so far. Co-founder Choy Travers said ads in markdown had not changed advertisers' results or win rate yet. The briefing does not report a longer-term figure.
What should I prioritise instead?
Crawlability, clean page structure and extractable claims on the pages you already have. The briefing gives no evidence that a format conversion lifts citations, while Google's own guidance is that markdown is not needed.
If a markdown mirror or an llms.txt rollout is sitting on your roadmap because a 2026 checklist told you to add it, it is worth rechecking what you expect it to deliver. Relevant Audience can look at where your pages actually get picked up by AI answers and put the effort where the evidence points.
