An engineering lead pings you with a two-week estimate for an llms.txt rollout. Meanwhile your answer blocks are missing on the pages that actually convert, and half your product pages have not touched dateModified in fourteen months. This is the actual choice practitioners face: twelve on-page techniques for AI citation, one engineering backlog, and a file format with no documented citation lift competing for the same sprint.
Key Takeaways
- Retrieval-layer techniques like answer blocks and dateModified freshness pay back faster than training-data techniques because models can act on them within a crawl cycle.
- llms.txt has no independent evidence of citation lift because no major model has confirmed it reads or weights the file.
- Sequence on-page work by how directly each technique feeds a model’s retrieval step, then move to entity and citation work once the retrieval layer is clean.
When Twelve Techniques Compete for the Same Sprint
Most lists of on-page AEO techniques read like a checklist: answer blocks, dateModified freshness, FAQ schema, entity disambiguation, canonical tags, semantic headers, table formatting, author attribution, internal linking for topic clusters, citation-ready summaries, structured data markup, content architecture. Treated as a checklist, all twelve look equally urgent. They are not.
Each technique feeds a different part of how a model finds and trusts a page. Some affect whether a page gets retrieved at all. Some affect whether, once retrieved, a model is willing to quote it. Confusing the two wastes engineering time on the wrong item first.
The AEO has two layers worth separating here. Retrieval presence is earned technically: structured, current content that a model can pull at query time. Training-data presence is earned slowly, through wide coverage across independent sources that eventually shapes what a model has absorbed. Answer blocks and dateModified sit in the first layer. llms.txt, if it does anything, would sit closer to the second, and even that is unproven.
Retrieval Presence Pays Back Faster Than Training-Data Presence
Answer blocks work because they hand a model something it can lift whole. A tight, self-contained paragraph that states a claim, a number and a source in three sentences is exactly the shape Perplexity and ChatGPT tend to quote inline. There is no ambiguity for the model to resolve, so there is nothing to lose in translation.
dateModified freshness works differently. Gemini cross-references Search, YouTube and Scholar, and freshness signals affect which version of a page it treats as current. A page with an honest, recently updated dateModified is telling every crawler that the content behind it has been checked, not just republished.
Neither of these requires a new file format, a new crawler directive, or a bet on how a model might behave in six months. They require someone to rewrite the top of a page and correct the timestamp. That is why they belong ahead of speculative infrastructure work in almost every prioritization conversation.
growth in AI brand presence for NewDay USA, tracked weekly across five engines over six months.See the case study
That growth came from consistent on-page and structural work across five engines, not from a single directive file. The lift was measured, not assumed, which is the standard every technique on the list should be held to before it earns engineering time.
Rank the Techniques by Mechanism, Not by List Position
A useful way to sequence the twelve is to ask, for each one, what specifically it changes in a model’s pipeline. Group them by mechanism and the order writes itself.
- Techniques that make a page quotable in isolation: answer blocks, citation-ready summaries, table formatting. These change what a model can lift without editing.
- Techniques that make a page trustworthy at crawl time: dateModified freshness, author attribution, canonical tags. These change whether a model treats the page as current and authoritative.
- Techniques that make an entity legible across the web: entity disambiguation, structured data markup, internal linking for topic clusters. These change how consistently a model can connect your brand to a claim, which is the mechanism behind trust as a graph of agreement rather than a score.
- Techniques that widen coverage over time: FAQ schema, semantic headers, content architecture. These compound. Once a model has cited a brand once, the next citation on a related query becomes more likely, and these techniques are what keep feeding that loop.
Notice llms.txt does not appear in any of the four groups. That is deliberate. It has not been shown to change retrieval, trust, entity legibility, or coverage for any of the major models.
llms.txt is infrastructure looking for a mechanism, and engineering time should follow proven mechanisms first.
What llms.txt Actually Does, and What It Does Not
llms.txt is a proposed convention: a plain-text file at the root of a domain that lists pages a site would like a model to read, similar in spirit to robots.txt. The idea is reasonable. The problem is adoption. None of the major model providers have confirmed they crawl it, weight it, or use it to decide what gets cited. ChatGPT pulls from across the open web. Claude leans on high-authority publications and documentation. Neither has published guidance saying llms.txt changes that behavior.
That does not make llms.txt worthless forever. It makes it speculative now. Building it today is a bet on future model behavior, not a response to current, observable citation lift. Compare that to answer blocks, which produce a visible change in whether Perplexity cites a page inline within a normal content cycle, no crawler cooperation required.
The distinction that matters here is Share of Citation versus Share of Voice. Share of Voice measures how often you are mentioned anywhere. Share of Citation measures how often a model actually names you as a source. llms.txt, even if it worked exactly as intended, would only ever influence which pages a model can see. It would not make the content on those pages more citable. That work still has to happen on the page itself.
A Sequencing Model You Can Defend to Engineering
When a lead asks why llms.txt is not sprint one, the answer is not “it does not matter.” The answer is that it has not been shown to matter yet, and eleven other techniques have a clearer path from change to measurable citation.
A defensible order looks like this: fix answer blocks and dateModified on pages that already rank for high-intent queries. Add structured data and entity markers where a brand name currently appears without disambiguation. Build out citation-ready summaries and FAQ schema on the pages driving the most organic traffic today. Only after that backlog clears does a directive file like llms.txt earn a slot, and even then, treat it as an experiment with a benchmark before and after, not a certainty.
This also keeps the work aligned with existing SEO and PR teams instead of routing around them. On-page techniques extend equity those teams have already built. Speculative crawler directives do not.
Where To Start This Week
Pull the ten pages generating the most organic sessions this quarter. Check each for three things: a self-contained answer block near the top, an honest dateModified, and clean entity markup on the brand name. Fix what is missing before any engineering ticket for llms.txt gets written. Run an AI visibility audit against those same pages to see which models currently cite them, then track that number weekly for six weeks before deciding whether infrastructure work is worth the sprint it would take.
FAQs
Should llms.txt Ever Be Built, or Skipped Entirely?
It should be built after the higher-certainty on-page techniques are in place, and treated as a monitored experiment rather than a guaranteed win, since no major model has confirmed it affects citation.
Which On-Page Technique Usually Shows Results Fastest?
Answer blocks tend to show the fastest change, because they affect what a model can quote at retrieval time without waiting for a training update or a new crawl cycle.
How Do I Know If dateModified Is Actually Helping?
Track whether pages with recently updated dateModified tags appear more often in AI Overviews or model citations compared to similar pages with stale timestamps, over a period of several weeks rather than days.
Does Fixing On-Page Techniques Replace the Need for Digital PR?
No. On-page techniques improve retrieval presence, while digital PR and citation building improve training-data presence over time. Both layers matter, and neither replaces the other.
