AEO Is the New IR Website Audit — and the ROI Numbers Are Vendor Fiction

A vendor will soon sell you “answer engine optimisation for IR,” and the pitch will come with a number. A 30-to-40-point gain. Forty hours saved a quarter. A 12% lift in valuation from better communication. The discipline they are selling is real and worth buying. The number attached to it is fiction. The hard part is holding both of those thoughts at the same time.

The first two pieces in this series argued that the machine is now your reader, and that the one thing it cannot reproduce is the human moment under live questioning. This piece is about the part the machine can read: your filings, your releases, your IR website. Making sure it reads them correctly, and from you, is a genuine new discipline. Believing the ROI a vendor quotes for it is how you end up paying for theatre.

The discipline is real

When an investor asks ChatGPT or Perplexity or Gemini about your company, the answer they get is assembled by a machine that has read your disclosure and decided what it says. Answer engine optimisation is the practice of making sure that answer is accurate, current, and sourced from you rather than from a third-party aggregator or a confident hallucination.

This is not vendor invention. In early 2025, Google, Microsoft, and OpenAI all confirmed that their AI features draw on schema.org structured data. The mechanism is documented and independent of anyone selling a service: structured, tagged content is read correctly, and unstructured content is read by guesswork.

The failure mode is concrete. XBRL International ran the test directly: it gave a language model a set of structured financial reports, and the model analysed them correctly. Given the identical numbers as a PDF, the model pulled a figure from the wrong column into a calculation that did not add up. Same data. Different format. The PDF is where your numbers go to be misread, silently, in the one place an investor’s machine is now looking.

So there is real work here, and most IR teams have not done it.

What the audit actually checks

Strip the jargon and an AEO audit is a short, unglamorous list.

Structure first. Is your financial and disclosure content available as tagged data and clean HTML, or is it locked inside PDFs? Where your filing regime offers structured formats, are you using them? A machine reads structure and misreads layout, so the format you publish in now decides whether it gets your figures right.

Source of truth next. When the machine answers a question about your company, is your IR site the authority it cites, or has something else taken that role? If your own site is not the machine-readable record, an aggregator or a stale, half-remembered figure becomes the answer instead. This is the one piece of ground IR cannot afford to cede, because everything downstream inherits whatever the machine treats as canonical.

Then recency. The AEO guides claim that most AI citations come from recently published content, often citing something like 95% within the first 45 days. That figure is itself a vendor number, so treat it with the same suspicion as the rest. But the direction is sound and independently sensible: timeliness is now an input to whether the machine sees you at all, not just good manners.

And finally substance, which is the part the playbooks underplay. A context-aware model rewards specific, verifiable facts and exposes vagueness. “We remain confident in our disciplined execution” is not just unpersuasive to a human; it is close to invisible to a machine, which has nothing concrete to extract. The durable move is the one good IR writing always required: anchor every reassurance to at least one verifiable fact and a sensible interpretation of it. That was sound advice before machines could read. It is now also how the machine finds you.

One caution while you are buying. Some of the audit checklist is genuinely established practice; some of it is not. The llms.txt file vendors increasingly insist on, for instance, is a proposal, not a standard. Do not let a real audit smuggle in speculative table stakes.

Now the fiction

Here is where the same vendor that sold you the real audit attaches a number that is not.

The performance claims circulating in the AI-for-IR market are striking and almost entirely unverifiable. Thirty-to-forty-point gains in AI visibility. Twenty to fifty hours saved per quarter. A 12.2% valuation upside attributed to communication. “99.9% of influence technologised away.” An “83% institutional AI embedding” rate. Every one of these is proprietary and self-reported, with no named third party and no published method behind it.

There is a clean test for any AI claim, and it sorts this market quickly. Is the number evidenced by someone with no stake in the sale, or supplied by the seller? Independent research in this field consistently quantifies the limits: one widely cited estimate puts the failure rate of generative-AI projects at up to 85%; the academic work on earnings calls measures the narrow human residual the machine cannot replicate; the quantitative research shows the edge from gaming a model decaying as the models improve. Vendors, just as consistently, quantify the upside. That asymmetry is the tell. A figure that only ever flatters the seller, with no method behind it, is not a metric. It is marketing.

This is not merely irritating. Repeating an invented AI-performance number is now a disclosure risk. The SEC’s first “AI-washing” actions in March 2024 fined two advisers a combined $400,000 for overstating their AI use; the first action against a public company followed in January 2025. NIRI has warned that AI-washing mirrors greenwashing, and that regulators want disclosure of real AI risks, not boilerplate. The gap between rhetoric and substance is already visible: a review of the 30 Dow companies found only one disclosed an AI audit. If you import a vendor’s fictional number into your own investor materials, you have not bought credibility. You have bought their exposure.

How to buy the discipline without buying the fiction

The governance scaffold already exists. The IR Society’s 2025 best-practice guidance is the most usable version: adopt a written IR AI policy with a permitted, restricted, and prohibited-uses matrix; keep documented human scrutiny of any AI output that reaches a public document; disclose material AI use in your reporting; and run real due diligence on vendors, including security and AI-management certifications and EU AI Act risk-tiering.

For the ROI claims specifically, one question does most of the work. Ask the vendor for the methodology behind any number they quote. If a named third party or a reproducible method appears, it is evidence and you can weigh it. If the answer is that the figure is proprietary, it is marketing, and you should buy the audit and discard the number.

Buy the audit. Ignore the number. The work is real even when the figure attached to it is invented.

And buy the audit you should, because the reason to do this work was never the valuation pop a vendor promises. It is that the downside is asymmetric. Being machine-unreadable does not leave you neutral; it hands the machine a worse source to answer from, and in the thin-coverage markets where the machine is often the only reader, there is no analyst layer to correct it. You do the audit to control what the machine says about you. That is reason enough, and it does not require a single invented statistic.

The discipline without the theatre

AEO is the new IR website audit. That sentence is true in both directions. It is a real, overdue, slightly boring discipline: structure your data, own your source of truth, publish on time, write in specific facts. Do it, ideally before a competitor’s cleaner machine-readability makes them the easier company for the machine to summarise.

It is also the newest surface for vendor theatre, and the numbers stapled to it are invented. Ignore them. The case for making your company machine-readable does not rest on a 12% valuation claim. It rests on the two arguments that opened this series: the machine is now the reader, and the human residual is the part it cannot fake. Between those two facts sits a simple obligation. The part the machine can read, make sure it reads correctly, and make sure it reads it from you.