Nyman Media's Signals bets that AI crawlers, not readers, are the audience to optimize for

Nyman Media launched Signals on Product Hunt with a one-line pitch: make your website work better for people and AI. That is the whole product description. No pricing tier, no benchmark, no named customer. Just a claim that a site can be tuned for two audiences at once, and an implicit bet that the second audience now matters as much as the first.

The bet is correct, even if the launch is thin. What is new here is not the idea of optimizing content for machines. That has been the SEO industry’s job for twenty-five years. What is new is that the machine reading your site is no longer a ranking algorithm deciding where to file a link. It is a model that will summarize your page, quote it, or skip it entirely, and the site owner has almost no visibility into which of those happened.

The shift nobody priced in

Search engine optimization assumed a stable deal. You structure your content, Google indexes it, Google sends traffic back. The crawler was a means to an end, and the end was a human clicking through. Signals is pitched at a world where the crawler is the end. An assistant reads the page and answers the user directly. The click never happens.

That changes the economics of a website in ways most operators have not internalized. If a model summarizes your reporting and the user never visits, your ad inventory is worth less, your subscription funnel leaks, and your analytics show a traffic decline you cannot explain. Publishers have been arguing about this for two years. The New York Times sued OpenAI and Microsoft in December 2023 over training data. News Corp signed a licensing deal with OpenAI in May 2024. Those are the visible fights. The invisible one is happening on millions of smaller sites that will never get a licensing offer and have no idea whether they are being read, cited, or ignored.

{/* TODO: verify the specific terms and current status of the News Corp–OpenAI licensing deal, and confirm the December 2023 filing date of the NYT suit, before publication */}

What “works for AI” actually means

Here is where the pitch gets slippery. Making a site work for people is a design problem with decades of accumulated practice: readable type, fast load, clear navigation, sensible hierarchy. Making a site work for a model is a different problem, and the industry has not settled what the answer is.

There are at least three competing theories. The first says structure matters most: clean semantic HTML, schema.org markup, explicit headings, machine-readable metadata. The second says it is about retrieval, not presentation, and that what matters is whether your content gets chunked and embedded well enough to surface in a retrieval-augmented pipeline. The third says the real lever is llms.txt, a proposed convention for telling models what they may and may not use, which is closer to a robots.txt for the generative era than to anything in classic SEO.

These theories imply different products. A schema-markup tool is not a retrieval-optimization tool is not a permissions-file generator. A launch that promises to serve “AI” without saying which layer it operates on is promising something the buyer cannot evaluate. Nyman Media may have a real answer. The Product Hunt listing does not state one.

The measurement problem is the real one

Even if Signals picks the right layer, the customer cannot tell whether it worked. Classic SEO had a feedback loop: rankings moved, traffic moved, you could attribute the change. Generative search breaks that loop. If a model cites your page in an answer, you often get no referrer, no impression count, no click. OpenAI, Google, and Perplexity have all shipped some form of citation or referral reporting, but the coverage is partial and the definitions differ across vendors.

{/* TODO: verify which AI search vendors currently expose citation or referral analytics, and what each reports, before publication */}

So a product that claims to improve your standing with AI systems is selling into a vacuum. The buyer pays, changes their site, and has no reliable way to confirm the change mattered. That is not a fatal flaw, and it is not unique to Signals. It is the condition of the entire generative-engine-optimization category right now. But it means the winners in this market will likely be the ones who also solve measurement, not just the ones who solve the markup.

Why the incumbents are exposed

The interesting structural point is who should be worried. Traditional SEO platforms built their businesses on a crawlable, rankable, attributable web. That web is being replaced by a probabilistic one where the same query returns different sources on different days and the source often does not know it was used. The tooling has not caught up. A startup that ships even a crude signal about how models are consuming a site is filling a gap the incumbents have been slow to acknowledge.

There is a second-order effect for AI builders, and it cuts the other way. If sites cannot measure their value to models, some will respond by blocking crawlers entirely. Reddit, Stack Overflow, and a long list of publishers have already tightened access or cut deals. Every site that walls off raises the cost of fresh training and retrieval data for everyone downstream. A tool that helps sites stay open by making their value legible is, in a small way, infrastructure for the model economy.

What to watch

The honest read on Signals is that the thesis is stronger than the launch. The claim that websites need to serve two audiences is right, and the market for helping them do it is real and mostly unbuilt. What is missing is the part that would make it a business: which technical layer it targets, how it measures success, and whether it can show a customer anything beyond a before-and-after screenshot.

Watch whether Nyman Media names a specific mechanism, publishes any benchmark against a named model or search product, or lands a customer willing to say on the record that the change moved a number. Any of those would turn a plausible pitch into a testable product. Until then, Signals is a correct diagnosis with the treatment left blank.