The Brave of
AI search.

Every AI search analytics vendor faces the same uncomfortable question: where does your conversation data come from? The honest answers across the industry range from "we make prompts up and replay them" to "we buy clickstream from brokers and prefer you don't ask." There is a better answer, and a browser company already proved it at scale. Brave built an independent search index almost entirely on data that users chose to contribute — anonymized, sanitized, and aggregated so aggressively that Brave itself cannot reconstruct anyone's session. We think that model is the right one for AI search too. This post explains Brave's Web Discovery Project, the analogy to what we do at Aiso, and why full opt-in isn't just the ethical choice — it's the more durable and more honest dataset.

What Brave actually built

In October 2021, Brave made its own search engine the default in the Brave browser and shipped something quieter alongside it: the Web Discovery Project. The pitch to users was simple. Brave Search wanted an independent index — results served from its own crawl-and-rank pipeline rather than white-labeled Google or Bing — and the fastest way to learn which pages matter is to learn from real browsing. But Brave is a privacy company, so the usual approach (silently hoover up everyone's history) was off the table. Instead, the Web Discovery Project is off by default. Users who opt in contribute search queries, result clicks, visited URLs, time on page, and some page metadata.

The interesting part is everything Brave does to make that contribution unlinkable to a person:

• Sanitization happens in the browser. Queries longer than 50 characters or 7 words are dropped. So is anything containing long numbers, emails, or credentials. Pages behind a login, pages flagged no-index, and URLs that look private (hashes, odd ports, IP addresses) never leave the device. Brave even fetches each candidate page a second time anonymously and compares; if the logged-out version differs, the page is treated as private and discarded.
• No user identifiers, ever. Records carry no UID. Messages are sent at randomized intervals in separate requests with one-time encryption keys, so they can't be correlated by timing. A proxy network hides the sender's IP.
• A URL needs a crowd before Brave can even read it. Under the STAR protocol (published at ACM CCS 2022), contributions arrive encrypted and only become decryptable once enough independent users have submitted the same URL. That's k-anonymity enforced by cryptography rather than by a privacy policy: a page one person visited stays unreadable, full stop.

The net effect, in Brave's own framing, is that reconstructing a user's session is not merely against policy — it is not technically possible. And the code is open source, so you don't have to take their word for it.

The analogy: conversations are the new clickstream

Swap "search index" for "AI search visibility" and the problem is identical. To know how AI assistants actually answer real people — which brands they recommend, which sources they cite, how answers shift week to week — you need real conversations, at scale, over time. The question is the same one Brave faced: how do you get that data without spying on anyone?

Our answer is the same as theirs: ask. Aiso's panel exists because users opt in, knowingly, with a clear value exchange — they get free access to premium AI models and features, and in return they agree to share their conversations. No silent SDK buried in a flashlight app, no resold extension telemetry, no scraping of pages people never meant to publish at scale. Conversations are anonymized before analysis, and what brands see is the aggregate: how a category is talked about across millions of real prompts, never an individual's chat log. We also publish exactly who our panel is and where it's skewed — sample composition, geography, intent mix, and the limitations — because a consented panel is the only kind you can afford to describe honestly.

The four ways the industry sources this data

To see why the opt-in model matters, line it up against the alternatives. Broadly, AI search and conversation intelligence data comes from four places today:

Synthetic replays have a real place — we use controlled replays ourselves when a question demands a fixed experiment, and we've written about how much run-to-run variance they carry. But replays answer "what does the model say to my prompt," not "what are real people actually asking, and what do they get back." For the second question there is no substitute for real conversations, and for real conversations there are only two paths: take them without asking, or ask. Brave demonstrated that asking scales.

Why opt-in data is also just better data

It would be easy to frame full opt-in purely as an ethics position. It is one, but it's also a quality and durability position, and that's the part most analyses miss.

1. You can describe a consented panel honestly

Because our panelists chose to join, we can study who they are and publish it — including the unflattering parts, like the fact that recruiting through Reddit ads skews the panel toward early adopters and technical users. Vendors sitting on gray-area data can't publish a demographics page without incriminating themselves, so their "market share" and "visibility" numbers arrive with no error bars and no account of sample bias. Transparency about limitations is a feature only consent can buy.

2. Consent survives crackdowns

The recent history of conversation data is a graveyard of gray-area sources: shared-conversation pages de-indexed overnight, browser-extension telemetry purged from stores, broker datasets pulled under regulatory pressure. Every product built on those sources inherited their shelf life. An opt-in panel has no such dependency — the data relationship is between us and people who said yes, and it renews itself for as long as the value exchange stays fair.

3. Aggregates are the product anyway

Brands don't need any individual's chat log; they need to know how assistants talk about their category across thousands of real prompts. Brave's insight was that if aggregate knowledge is the goal, you can make individual knowledge technically unreachable and lose nothing. The same holds for AI search: anonymized, aggregated conversations contain all the signal — which brands get recommended, in which intents, what people are actually asking, how that shifts over time — with none of the surveillance.

What this means if you're evaluating AI search tools

Whether or not you ever use Aiso, put this question to every vendor whose dashboard claims to know what people ask ChatGPT: "Did the humans in your dataset agree to be in it?" Then three follow-ups:

• Provenance: Is the data synthetic, scraped, purchased, or opt-in? If the answer is vague ("proprietary panel partnerships"), assume the least flattering option.
• Transparency: Does the vendor publish the panel's size, geography, and skews? A vendor that can't show you its sample can't defend its numbers.
• Durability: If one platform changed its terms tomorrow, does the dataset survive? Consented panels do; harvested corpora historically have not.

Brave asked these questions of search a decade after everyone assumed surveillance was the price of a good index, and built a real alternative. AI search is young enough to get this right from the start. That's the project we're running: consent-first conversation data, anonymized and aggregated, with the panel's shape published for anyone to audit. The Brave of AI search isn't a slogan — it's a sourcing policy.

FAQ

Is Aiso affiliated with Brave?

No. Brave is the inspiration and the proof of concept, not a partner. We admire the Web Discovery Project's design and hold our own collection to its principles: strict opt-in, clear value exchange, anonymization before analysis, and aggregate-only reporting.

Can I see what Aiso's panel looks like?

Yes — we keep a standing, regularly refreshed write-up of our panel's demographics, geography, intent mix, and limitations, including the skews introduced by how we recruit.

Does opt-in data cover enough volume to be useful?

Brave Search answers billions of queries a year on an index built largely from opt-in contributions. Our panel spans 5M+ unique IPs and millions of conversations across ChatGPT, Gemini, and Claude in 120+ countries. Consent and scale are not in tension; they compound, because a fair value exchange keeps the panel growing.