A/B test what AI says about you.

The customer journey now happens before the visit: people ask an AI agent to compare, recommend and decide. Oomira asks 4 named agents (Claude, ChatGPT, Perplexity and Gemini) the questions buyers, investors, candidates, partners and press actually ask about a company, cold, with live web search and nothing supplied by us. It shows each answer word for word, marks its claims for you, mixed or against you, and traces the answer to the pages it opened: the company’s own, its competitors’ and the outside sites that shaped it. Then it writes a hypothesis about what is driving the answer, and when a page changes it asks the same question again and compares with the answers before.

Web analytics measures what people do once they reach a site. SEO and technical tools measure whether pages are indexed, reachable and structured. AEO and visibility tools measure whether a company appears in AI answers and its share against competitors. Oomira answers a different question: does changing this page change what the agent says. It works in tests: controlled before and after checks on the pages and sources shaping an answer. Not a split test, since there is one web and one set of pages, so the control is holding everything else still: the same question, the same agents, asked cold with nothing supplied by us, before a change and after it. They are complementary: Oomira finds the page and runs the test, and the other tools show the traffic move.

Four answers, word for word, with every sentence tagged: marked for you, mixed or against you, and carrying the citation that agent reported for it. That is every sentence of every answer, on every plan. Every page the agents opened is fetched and kept, yours and other people’s, and searched when Oomira forms and traces what it concludes, so a finding rests on the pages that produced the answer rather than on a summary of them. Out of that comes a headline, which says what the answers come to. A test is designed from those answers when you ask for one: a change, a prediction, and the question to ask again. Your free first question comes with its first test. On a plan, from the third asked question, Oomira also writes the Beliefs the agents hold across the account’s questions, and updates them when you ask. Judging each sentence against the text of those pages, rather than against the agent’s own account of what it read, is the Enterprise claim analysis.

Your own assistant holds your history, your earlier questions and often your own website already in the conversation, so it answers about a company it has been introduced to. Oomira asks four agents cold: a blank slate, nothing supplied by us, one live search each. That is the answer a stranger gets, and it comes back with every page each agent opened, so a claim can be traced to the page that caused it rather than argued with.

As fast as the agents fetch the page again: on MycoFutures' test, Claude cited the sources they added five minutes after the change went live, while the other agents had not yet opened it. Every question is an evidence environment: each search the agent ran, each page it opened and each page it cited is recorded, every sentence carries the citation the agent gave for it, and the pages themselves are kept and searched when Oomira works out what produced the answer. So a claim is either backed by something that agent just read, in which case the page that caused it is named and can be changed, or it is backed by none of them, which means it came from the model or from something cached, and that is a different problem with a different fix. Change the page or the outside source, ask the same question again, and compare with the answers before. Answers vary between days, so a test predicts that fewer agents will say the thing rather than that one agent will say a given sentence.

About a minute. Your first question is free and needs no account: it goes to 4 agents, and two of the answers show at once. With a work email the rest open, and the full report is written about three minutes after the address: all 4 answers, the pages behind them, and your first test. After that nothing runs on a clock. A question, a test or a recheck goes when you press.

Sometimes the very next time it is asked, sometimes not. An agent that opens the changed page can pick it up immediately; a claim that came from an outside site, or from something cached, moves when that source moves. That is why a change is a test rather than a fix: a recheck asks the same question of the same agents and compares with the answers before, instead of waiting and hoping.

Your first question is free and needs no account. A work email adds all 4 answers, the full report, your first test and one free recheck once your change is live. Another question, with its test and 3 rechecks, is $29, one payment, with no plan. The question goes cold to all four agents with every sentence traced to the page the agent cited for it, and its test is designed when you press for one. Each recheck asks all four again after a change: up to 16 agent answers on the one question. Your free report and its test are saved to an account, with 3 rechecks of its own. You choose the question after you pay, and it is kept until you use it. Pro is $99 a month for one company: 10 questions and 10 tests a month, 3 rechecks per question, each counted separately. Asking a question puts it cold to all four agents and traces every sentence of every answer to the page the agent cited for it. A test is designed when you press for it on a question: the change for your site, the page, the wording and the prediction, from that question's answers. A recheck asks the same question again after a change, up to 3 times per question a month. Nothing runs unless you press it. Redrawing a test from everything your account has learned since uses one of the month’s tests. Site checks are not counted, and Beliefs are written from your third asked question and updated when you ask for it. 1 user. Teams is $259 a month: the same questions, tests and rechecks, for 5 users. Enterprise is by contract and adds the fact check, the claim analysis, follow-up questions, site checks across subdomains, and several brands or markets. A month’s questions, tests and rechecks do not carry over.