Most of what is sold as AI search optimization is unproven, and some of it is measurably useless. Google's documentation says you need no special files or schema to appear in its AI features, and an Ahrefs study of 137,210 domains found 97 percent of llms.txt files went unread in May 2026. What decides whether an assistant can cite you is older and duller. Your content has to be in the HTML, and your claims have to be checkable.
What actually changes when a buyer asks an assistant instead of typing a query?
The buyer stops reading your page and starts reading a paragraph about you.
That is the entire shift. A search box returns a list of destinations and asks the buyer to choose. An assistant returns one synthesized answer and does the choosing, citing a few sources along the way. Your page stops being the destination and becomes raw material for a paragraph somebody else composes.
Three things follow, and only the third is interesting. The click may never happen. Position matters less than whether your sentence is liftable. And the assistant has to be able to state a claim about you and attach your name to it. That last one is a different requirement than ranking. Ranking rewards a page. Citation rewards a sentence.
So the practical question stops being "where do we rank for custom software development" and becomes "if an assistant is asked who builds multi-tenant SaaS platforms in Las Vegas, is there a passage on our site that answers it in one paragraph, with our name attached clearly enough to survive being summarized."
Why can a site that reads well to a human still be uncitable?
Because a person renders JavaScript and a retrieval bot often does not.
This is the least glamorous failure and the most common one. If a marketing site ships an empty container and paints the content in after load, a human sees a polished page and a crawler sees nothing at all.
The most concrete public measurement of this comes from a scan of the Y Combinator Spring 2026 batch published on June 22, 2026, which fetched all 195 reachable company sites. 164 served real content directly in their HTML. 17 of them, about 1 in 11, were empty shells. The same scan found that 50 percent carried any structured data at all, that FAQ markup appeared on 19 percent, and that 9 percent blocked at least one major AI crawler. That is one scan by a small tool vendor rather than a peer-reviewed study, so hold the exact figures loosely. The direction is still worth sitting with: a meaningful slice of well-funded, technically competent startups had marketing sites that said nothing to a machine.
Those are not optimization problems. They are plumbing problems, and they are invisible from the inside because the site looks fine in a browser. It is the same shape of failure described in why executive dashboards fail: the expensive breakage happens upstream of the thing everyone is looking at.
What the evidence says about the files being sold to you
This subject deserves bluntness, because it attracts more unfounded marketing than nearly any other topic in B2B software.
On llms.txt: Ahrefs published an analysis on June 15, 2026 covering 137,210 domains. 38,360 of them (28 percent) published a valid file. Of those, in the study's words, "97% of those files received zero traffic in May 2026." Of the requests that did arrive, only 1.1 percent came from AI retrieval bots. Most came from SEO audit tools and profilers checking whether the file exists. The most telling finding: "Zero requests came from AI bots for llms.txt files that don't exist. They never go looking." Google's John Mueller has described the format as a "temporary crutch, perhaps to save some tokens" for AI coding tools reading developer documentation.
Commerce Beacon publishes an llms.txt. It would be dishonest to imply it is doing work it is not doing. It costs nothing to maintain and it is a cheap bet on a format that may matter later. No published evidence shows that it affects whether an assistant cites us.
On schema and special markup: Google's documentation for its AI features, last updated December 10, 2025, is unusually direct. "You don't need to create new machine readable files, AI text files, or markup to appear in these features." On structured data it is equally plain, stating that there is no special markup from the schema vocabulary that you need to add either. Eligibility is the ordinary bar, meaning the page must be indexed and eligible to be shown with a snippet. "There are no additional technical requirements."
On FAQ markup specifically: Google restricted FAQ rich results to well-known government and health sites in 2023, and on May 7, 2026 they stopped appearing in Google Search entirely. Google's own guidance is that "You can remove the FAQ structured data from your code, if you want but you can also leave it," noting that "Other search engines may be able to continue to process it and use it for their own purposes."
This site emits FAQPage schema on its articles and service pages, alongside Organization, WebSite, and Person markup in a single linked graph. Since the May 2026 change, that FAQ markup earns no rich results in Google at all. We keep it because it costs nothing, because other consumers still parse it, and because the discipline is the real return. Writing a genuine FAQ forces you to state a question plainly and answer it in one self-contained paragraph. That is the citable-passage exercise. The schema is a byproduct of doing it.
Notice the pattern. The tactics that are easy to sell, a file or a markup block, are the ones with no evidence behind them. The thing that works is the thing that was always work.
So what actually seems to matter?
The durable mechanics, in rough order of return:
- Ship content in HTML. If a crawler cannot read it without executing JavaScript, nothing further on this list matters.
- Answer the question in the first paragraph. The liftable unit is a paragraph that still makes sense with everything around it removed.
- Be clear about who you are. Entity clarity is mostly not a schema problem. It is whether your name, your location, what you build, and who runs the company are stated consistently in ordinary text across the site.
- Make claims checkable. An assistant repeating a claim about you is putting its own credibility behind it. Specific, dated, verifiable statements survive summarization. Adjectives do not.
- Do not block the crawlers you want to be read by, and confirm that rather than assuming it. None of that is new. It is the same instruction set as writing a clear page for a human who is in a hurry, which is why the honest version of this advice is less marketable than a file you can generate in an afternoon.
Here is the part most of the category will not say plainly: nobody has published a credible dataset on what share of B2B software buyers actually shortlist vendors through an assistant, or on how much a citation is worth compared to a click. The widely circulated figures on AI referral traffic and click loss mostly trace back to single vendor studies or to nothing identifiable at all. Anyone quoting a precise percentage should be asked for the primary source, and the request will usually end the conversation.
What is defensible is narrower. Assistant behavior changes constantly, so building for any specific engine's current quirks is a poor investment. The mechanics underneath, being crawlable, being clear, being checkable, have been stable for twenty years and are the same mechanics that make a site work for a person. That is where Commerce Beacon puts the effort when building AI sales funnel systems or a SaaS platform: instrument what you can actually measure, be honest about what you cannot, and do not buy the file.