Tokens Became a Commodity in 18 Months. AI Search Is Next.

Three years ago, access to a frontier language model looked like a moat. Today a token is a commodity that trades on price, latency, and context window. The numbers are almost hard to believe. According to the Stanford AI Index, the cost of querying a model at GPT-3.5-level performance fell from $20 per million tokens in November 2022 to $0.07 by October 2024. That's a 280-fold drop in about 18 months. Epoch AI's pricing analysis puts the annual decline at anywhere from 9x to 900x depending on the capability milestone. And the quality gap between providers keeps shrinking: open-weight models now trail the closed frontier by a matter of months, not years.

Everyone who bet that owning the model was the business is finding out that the model was an input to the business. Inputs get commoditized. Margins go to whoever owns the customer.

AI search is running the same play on a compressed timeline, and most of the people building in the space are not pricing that in.

Lightning in a bottle, or a bottle factory

The current narrative around AI search is that somebody is about to catch lightning in a bottle. Be the next Google. Own the front door of the internet.

But Google was never just a better search engine. It was two decades of compounding defaults and habit. Court documents in the DOJ antitrust case showed Google pays roughly $20 billion a year to Apple for default placement in Safari alone, part of more than $26 billion in annual default deals. Nobody spends a quarter of their operating income on distribution if the product wins on quality by itself. The moat was placement and muscle memory, not technology.

AI search has none of that working in its favor. The underlying capability is rented from the same handful of foundation models, the same commoditizing token layer described above. The answers themselves converge: ask five answer engines the same question and you get five paraphrases of the same synthesis, pulled from the same crawled web. And switching costs round to zero. Nobody's email lives in their answer engine. There is no file storage, no social graph, no twenty years of habit. Users have already absorbed the meta-lesson that these tools are interchangeable, because they watched the model layer become interchangeable first.

A commodity input producing a convergent output, sold to users who can leave in one tab, is a race to the bottom. The only open question is speed.

Then the ads show up, and the clock accelerates

The bull case for monetizing AI search is "Google's ad business, but conversational." What's actually unfolding looks more like Google's fifteen-year decline as a trusted answer machine, run in time-lapse.

The ads are no longer hypothetical. OpenAI began testing ads in ChatGPT in early 2026 on its Free and Go tiers, with entry pricing reportedly starting around $200,000 and early partners like Best Buy and Expedia. The pilot reportedly passed $100 million in annualized revenue within two months at roughly $60 CPMs, about three times typical Meta pricing. Google has pushed ads into AI Overviews and AI Mode, and by some tracking, sponsored placements now appear alongside roughly a quarter of Google's AI answers, up from around 5% a year earlier.

The most honest data point in the whole category might be the company that quit. Perplexity launched its ad product in late 2024, then pulled the plug in early 2026, with executives telling the Financial Times that ads risked making users "suspicious of everything." That's the bind, stated out loud by an insider. A single synthesized answer is the entire product. There is no below-the-fold to hide the ads in, no organic listing sitting next to the sponsored one for comparison. Once money can touch the answer, the user has to wonder about every answer.

Google could degrade slowly because it had lock-in. It took a decade and a half of ads colonizing the results page before users started drifting, and even now the erosion mostly shows up as behavior inside Google's own walls: Pew found that when an AI summary appears, users click through to a website only 8% of the time, versus 15% without one, and only 1% ever click a cited source. The challengers get to run the same trust-burning playbook without the lock-in that made it survivable. Race to the bottom, but faster, because the bottom is closer and the brakes were never installed.

What this means for brand

If you accept the premise, the strategic implications invert most of the current playbook.

Stop treating AI search as a channel to be won and start treating it as a compression algorithm to be survived. In classic search a brand had real estate: a listing, a snippet, a page-one presence. In AI search your brand gets compressed into a clause. "For running shoes, most people recommend ___." You no longer occupy a position on a results page. You occupy, or fail to occupy, a position in the model's prior. And that position isn't bought at auction. It accumulates upstream, in training data, citations, reviews, and cultural presence that existed before the question was ever asked.

Put more bluntly: brand becomes the prior. When the interface is a synthesized answer, the names that get mentioned are the ones with overwhelming category association, the ones the corpus of human writing reaches for by default. The unfashionable fundamentals (distinctiveness, mental availability, category entry points) turn out to be the durable inputs to AI visibility. You cannot SEO your way into a language model's instincts. You can only be famous your way in.

Meanwhile, paid AI placement is shaping up to be a worse deal than search ads ever were. One answer slot instead of ten links. Entry minimums in the six figures. CPMs at three times Meta's rates, inside products whose users grow more skeptical with each sponsored synthesis, on platforms that are commoditizing against each other. Brands that over-index on these placements will be funding the degradation of the channel while the thing that actually drives being named in the answer, organic mindshare, quietly atrophies. The Pew click data already shows what happens to visibility that depends on someone else's interface.

The moat migrates to demand, not discovery. In a world of commoditizing intermediaries, the brands that win are the ones people ask for by name. "Is [Brand] good for X" instead of "What's good for X." Navigational demand is the one query type no answer engine can intermediate away, and building it looks like brand marketing, community, product distinctiveness, and cultural presence. It looks, inconveniently, like the long game.

The bet, restated

The token bet was that owning intelligence would be the business. It wasn't. Intelligence became the commodity, and value moved to whoever owned the customer relationship on top of it.

The AI search bet is that owning the answer will be the business. It won't be. Answers are already converging into a commodity, ad load is accelerating the trust collapse, and value will move, again, to whoever owns the relationship that exists before the question gets asked.

That relationship has a name. It's called a brand. And in a race to the bottom, it's the only thing that doesn't have to run.