The Google AI Moat Nobody Is Talking About

For two years the AI debate has been fought on the wrong battlefield: open weights versus closed, Llama versus GPT, whether DeepSeek’s latest release means nobody has a moat. Marc Andreessen says models are commoditizing, and the labs’ price wars suggest he’s right. But the actual moat was never in the weights. It’s a database Google started building in 1998 and has never stopped building: the search index. The AI transition didn’t make it obsolete. It promoted it.

From product to substrate

Search used to be a destination. You typed, you clicked a blue link, Google sold an ad against your attention. Now it’s a correctness input for AI models. A huge class of ordinary prompts can’t be answered from model weights at all — is so-and-so still the CEO, what does this cost, did the bill pass, who won last night. Without live retrieval the model isn’t slightly degraded on these; it’s confidently wrong about anything that moved after its training cutoff. That’s why every serious chatbot now searches the web before answering, silently, sometimes several times per prompt.

Those “several times” matter. Agentic workflows fan out — one question can trigger a dozen retrievals as the model decomposes it, cross-checks, and refines. On current Gemini models, Google bills grounding per search the model decides to run, not per prompt you type. Multiply across RAG pipelines, research agents, and coding assistants, and search stops being a feature. It’s the oxygen supply. So the question that matters isn’t whose model is best. It’s who owns the retrieval layer every model now depends on.

The physics of the index

Google dominates the index layer. The only other company that maintains a full-scale Western index today — Microsoft — holds roughly a tenth of Google’s scale. Everyone else rents, and has for a generation: DuckDuckGo never had an index of its own, Yahoo abandoned its own in 2009, and nearly every other “alternative” engine you’ve heard of syndicated Bing’s or Google’s. The retail search market wore a dozen masks over two wholesalers, one of them ten times the other.

And look at what even distant second place cost. Microsoft spent two decades and billions on Bing — with Windows distribution, Azure infrastructure, and an OpenAI partnership — and got an index in the low tens of billions of pages against Google’s hundreds of billions spanning more than a hundred petabytes. Twenty years and near-unlimited capital bought a tenth of the incumbent’s scale. That is the entry price for junior membership in the club. Nobody else has paid it, and nobody is trying — because the ladder has two cliffs, not one. Microsoft is a tenth of Google, but it still holds what no one below it has: decades of real consumer query traffic through Bing, Windows, and Edge feeding its ranking signals, and crawl infrastructure of a maturity no independent can match. Brave’s pages don’t come with twenty years of behavioral data. Second place is a distant second and still unreachable.

The gap never closed because an index isn’t built; it’s maintained. The moat is the perpetual recrawl: hundreds of billions of pages refreshed continuously, news revisited in minutes, a spam-filtering apparatus hardened by twenty-five years of war with the SEO industry, ranking signals distilled from trillions of human queries. And the index doesn’t sit alone — it’s fused to the Knowledge Graph, Maps, Shopping, YouTube, and the telemetry of Chrome and Android: a continuously refreshed model of the world. You don’t catch up to that with capital. You’d have had to be doing it since 2000.

Then, in 2025, the junior partner stopped wholesaling. Microsoft killed the open Bing Search API and pointed developers at “Grounding with Bing Search” inside the Azure Agents stack: $35 per thousand transactions, a steep increase over the old API for most workloads, and no longer a search API at all — raw results are never exposed, only a Microsoft model’s synthesis of them. Microsoft didn’t give up on search; it gave up on selling it. It adopted the dominant player’s playbook instead: never rent out the eyes raw — only rent your own brain wearing them, at a premium. When the distant second copies the leader’s enclosure strategy rather than undercutting it, the market isn’t being contested. It’s being farmed — one big farm and one small one, in its shadow.

The obvious rejoinder: maybe AI grounding doesn’t need Google-grade coverage — a curated, high-signal slice of the web plus a good model is “good enough.” But the general-assistant business lives in the tail curation excludes: the store hours, the spec sheet, the county ordinance, the forum thread where somebody already solved your exact problem. And the market has voted. The labs pay the toll today, and the ones rich enough to attempt independence — Perplexity with its partial stack, OpenAI with its reported in-house crawl — have spent fortunes without escaping it. If good enough were good enough, somebody with a hundred billion dollars would be acting like it.

Twelve months of receipts

August 2025: Bing’s API dies. January 2026: Google announces its own cheap Custom Search JSON API — $5 per thousand, the indie developer’s workhorse for nearly two decades — will return HTTP 410 Gone on January 1, 2027, steering everyone toward enterprise-priced Vertex AI. February 2026: Brave, now the only independent Western index at scale, kills its free tier and meters every developer at $5 per thousand. Brave’s own executives said it plainly: Bing’s exit made them “the only independent search API in the market at scale.”

The rate card

What it costs to give an AI model eyes, per 1,000 searches, July 2026:

Google — Gemini grounding (3.x) — $14
Google’s index, bundled to Gemini. Billed per search the model fires, not per prompt; model tokens extra.

Google — Gemini grounding (2.5) — $35
Same index, per-prompt billing.

Google — Custom Search JSON — $5
Raw Google results. Closed to new users; dead January 1, 2027.

OpenAI — web search tool — $10
Undisclosed index. Content tokens billed on top; some cheaper models charged a flat 8,000 input tokens per search.

Anthropic — web search tool — $10
Reportedly Brave’s index. Results billed as input tokens in that turn and every later turn.

Brave — $5
The last independent index (~30–40B pages). Free tier killed February 2026.

Exa — $7
Neural/semantic index. Smaller, curated coverage.

Tavily — $8–16
LLM-cleaned content over rented retrieval. Credit-based.

Microsoft — Grounding with Bing (Azure) — $35
Bing’s index, bundled to a Microsoft agent. Raw results never exposed; requires the Azure Agents stack; billed per tool invocation, and the model can invoke it multiple times per run.

Microsoft — Bing Search API — dead
Raw API decommissioned August 2025.

Two things jump out.

First, these prices are the entry toll, not the total. Retrieved content bills again as model tokens — often a multiple of the search fee itself. OpenAI charges some of its cheaper models a flat 8,000-token block per search. Anthropic’s results re-bill as input on every later turn of the conversation. On Gemini, the model decides how many searches to fire, invisibly, and you pay for each one. A modest product handling a million retrieval-backed prompts a month, at two searches per prompt, pays $20,000–$70,000 a month in search fees at the majors’ rates before a single token. Perplexity handled roughly 780 million queries in May 2025 — run that volume through any rate card and you understand why it spent a fortune building partial infrastructure of its own, and why almost nobody else can.

And the toll’s share is growing. Token prices collapsed an order of magnitude in two years — flagships repriced from $15 to $5 per million input tokens, budget models selling input at twenty cents. The search toll hasn’t moved: OpenAI and Anthropic at $10 since launch, Brave from free to metered, Bing from metered to enclosed. On a cheap model the search fee already exceeds the token cost of processing the results, and every model price cut shifts the ratio further. Inference is deflating like compute. The toll is priced like real estate.

Which hands Google a structural advantage at the model layer that has nothing to do with Gemini’s quality. Every rival’s live-world answer carries a real cash cost — a search fee paid to somebody. Google’s carries almost none: the index is already built, already running, already paid for by the ads business, and one more query against it costs a rounding error. As tokens race toward zero, the toll becomes the dominant marginal cost of a grounded answer — and Google is the only company that doesn’t pay it. In the endgame where models are commodities, every live-world answer costs Google nearly nothing and costs everyone else a fee that Google influences. You don’t need the best model to win that game. You need the cost floor, and only one company has it.

And the cost floor comes with a quality ceiling for everyone else. Nothing obliges Google to sell its best retrieval. What Gemini uses internally can be the whole crown jewels — the full ranking stack, the Knowledge Graph fusion, the freshness tiers, query understanding refined on a quarter century of traffic — while the grounding API exposes whatever slice Google chooses. Rivals can’t verify the gap, can’t buy the difference anywhere else, and can’t build it. And the incentive runs one way: every point of retrieval quality held back from the API is a point of advantage handed to Gemini. The toll road sells everyone else the service road and keeps the express lane for itself.

Second — my opinion, but the table argues it for me — these are monopoly prices. Google serves more than five trillion consumer searches a year for free, by its own disclosure, so the marginal cost of a query is a rounding error against these rates. Yet it prices grounding at $14–$35 per thousand, and the market arranges itself underneath: OpenAI and Anthropic, rivals who agree on nothing, both at exactly $10. Anthropic’s tool is reportedly Brave underneath — a clean 100% markup on a $5 input. Brave sits at $5 because with Bing gone and Google’s cheap tier dying there is nothing below it to fear. In a competitive infrastructure market, prices converge toward cost. Here they converge toward whatever Google charges, minus a courtesy discount. That’s a price umbrella, and the company holding it owns the index everyone else is approximating.

And the umbrella never gets tested, because nobody shops across it. At equal prices, everyone would choose Google’s index — it’s simply the best. But Google is also a competitor at the model layer, and that poisons the transaction twice over. A lab’s query stream is competitive intelligence of the highest grade — a live feed of what your users ask, what your agents are working on, where your model fails. And the queries are more than intelligence: they’re fuel. Google’s ranking quality was built on decades of human query traffic; chatbot traffic would hand it the equivalent signal for the AI era — how agents phrase requests, what they fan out into, which pages satisfy them. Route your retrieval through Google and you leak your product while paying to deepen the very moat you’re renting. So search is only sold bundled to each owner’s own models, no neutral market ever formed to discipline the prices, and the cleanest evidence is on the table above: Google is an investor in Anthropic and one of its cloud providers, and Anthropic still reportedly runs its search through Brave’s thirty-billion-page independent index rather than its own backer’s. When companies pay to not use the best product, the duopoly doesn’t need to collude. The structure of who-sees-whose-queries does it for them. And it explains the strangest number on the table: Microsoft charging Google’s $35 for an index a tenth the size. In a market where quality set prices, that would be laughable. In a market where half the customers can’t touch Google at any price — because Google is their competitor and their queries are its fuel — it’s simply what captivity costs. Microsoft isn’t pricing against Google. It’s pricing for Google’s refugees, and the $35 is its bet, written on a rate card: that the AI industry would rather pay a premium for the smaller index than hand its work to Google.

The stories everyone is talking about

Three adjacent stories get real attention. All three confirm the thesis.

Cloudflare has spent a year building tollbooths against AI crawlers — default blocks, a pay-per-crawl marketplace, now a “Pay Per Use” scheme — and said the quiet part loudly: Google’s crawler sees roughly twice the information available to any leading AI company, because Googlebot feeds search and AI through one pipe and no publisher will block it. They don’t want to. Being in Google’s index is where the traffic comes from, and an entire industry — SEO — consists of publishers doing free labor to be crawled better: sitemaps, schema markup, pages tuned to Google’s specifications. Every other AI crawler is an intruder to be blocked or billed. Google’s is a guest the web dresses up for. Nobody submits a sitemap to OpenAI. Even if regulators split Googlebot in two tomorrow, the search half keeps twenty-five years of index, query data, and freshness infrastructure that no ruling can redistribute — an index the world’s publishers improve daily, for free.

The licensing wave: as publishers wall off content, the premium head of the web is being carved up in direct deals — Axel Springer and Condé Nast with OpenAI, Reddit with Google at a reported $60 million a year. Some read this as a route around the crawled index. Two problems. Licensing covers the head, not the tail where the ordinary prompts live — no consortium of publisher deals covers the store hours and the spec sheets. And the bidding war is among companies whose content Google’s crawler still largely reads through the front door. Google bids when it wants to and abstains when it doesn’t. The closing of the web is an expense for everyone else and an option for Google.

The antitrust case: Judge Mehta’s September 2025 remedy declined to break off Chrome or Android, barred exclusive default deals while allowing paid placement, and ordered some sharing of index and behavioral data. Alphabet’s stock jumped eight percent on the ruling. The data-sharing provision hands competitors a snapshot, not a pipeline. The value is the machine that keeps the copy current, and the machine wasn’t on the table.

Free brain, no eyes

The open-source movement won a real victory: anyone can download a frontier-class brain, fine-tune it, run it on their own hardware. Model capability — the thing everyone spent two years treating as the moat — genuinely commoditized.

But a brain describing a world it stopped observing at its training cutoff is a brain in a jar. To be right about anything alive, it needs eyes — and the eyes were never open. The best pair belongs to a database in Mountain View that took a quarter century to build, that the only company to attempt a rival gave up matching and now farms alongside it, and that regulators declined to touch. Every AI company on earth now rents eyes by the query — from Google, from its junior partner, or from the one independent left — at rates set under Google’s umbrella.

Open weights got you a free brain. Nobody is handing out free eyes — and only one company owns a pair that sees the whole thing.