Which AI answer engines actually matter?

The question usually arrives attached to a spreadsheet. Fifteen rows, a column headed optimization tactics, and a quiet suspicion that no team has time to do fifteen things. Somewhere on that list is usually an engine that shut down in January.

The spreadsheet is the wrong shape. Most answer engines are not search engines. They are a language model sitting on top of somebody else's retrieval layer, and there are far fewer retrieval layers than there are logos. Sort the market by index instead of brand and fifteen targets collapse into about four.

Start with the indexes, not the engines

An answer engine needs a model to write the answer and a supply of fresh web documents to write it from. Training a competitive model is expensive but tractable. Building a web-scale index is brutally expensive, so almost everyone rents one.

The Apple row is the newest. Apple confirmed at WWDC 2026 that its next-generation Siri runs on Google Gemini, under a deal first reported in January 2026. Siri did not become a new target; it moved into an existing column.

The Copilot row is the best documented. Microsoft's Copilot Studio guidance on generative answers from public websites spells out the pipeline: the query becomes a Bing Custom Search query, the top results are collated and grounding-checked, then summarized. Bingbot does the crawling, and there is no separate Copilot crawler.

Two share numbers that disagree, and the gap is the finding

Two credible public datasets exist on AI engine share, and they disagree. Either one alone will send you to the wrong place.

StatCounter measures referral traffic: clicks that leave a chatbot and land on a third-party website. First Page Sage measures US usage. An engine can be used constantly and almost never send anyone anywhere.

That divergence is the finding. Gemini at 8.65% of referrals but 27.7% of usage is far bigger as an answer surface than as a traffic source; Claude at 2.91% against 10.3% has the same shape. Perplexity is the mirror image — 7.07% of referrals on 2.0% of usage, because it links out on nearly every answer. Judge these engines by referral sessions alone and you will underrate the two growing fastest.

The engines that carry real weight

ChatGPT remains the centre of gravity on both measures — 78.16% of referrals, 52.7% of US usage — though both figures are falling. It cites with clickable links and tags outbound clicks with utm_source=chatgpt.com, the easiest citation proof any engine offers. Its retrieval leans on a Bing-derived layer, and visibility is governed by OAI-SearchBot, not GPTBot.

Google's surfaces are one target wearing four faces. AI Overviews, AI Mode, Gemini, and now Siri all draw on the same index. Alphabet's Q4 2025 earnings put AI Mode at 75 million users by December 2025, and Google said at I/O 2026 that AI Overviews reach 2.5 billion monthly users. All of them cite, and Google's documentation is blunt about what eligibility takes.

To be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search.

Google Search Central, AI features documentation

Claude cites its sources and has run web search on Brave's independent index since March 2025. Brave's index is genuinely separate — over 30 billion pages, roughly 100 million daily updates — making Claude the one major engine you cannot reach through Google or Bing. One third-party vendor study found Claude's cited results matched Brave's top organic results 86.7% of the time, suggesting it re-ranks Brave's output comparatively little. Treat that as indicative rather than established.

Perplexity is the most citation-dense engine on the list, and the one whose retrieval is least settled. It operates a documented crawler, PerplexityBot, which its docs describe as designed to surface and link websites. One operator analysis argues the crawler is close to a red herring for live answering, and that Perplexity queries Google and Bing search APIs at query time — a third-party claim, not documentation.

Microsoft Copilot takes 3.19% of referrals and 1.3% of US usage, which understates it — much of its footprint sits inside Microsoft 365, where nobody clicks out at all. It cites, it runs on Bing, and the optimization surface is doing Bing Webmaster Tools properly: sitemap submitted, IndexNow enabled, crawl errors cleared.

Yahoo Scout, the large launch nobody is optimizing for

Scout is the entrant most readers have never heard of, and the strangest omission in 2026 roundups. Yahoo launched it in beta on January 27, 2026 across roughly 250 million US users spanning Mail, News, Finance, Sports, Search, and Shopping — distribution most startups would trade a funding round for, switched on at once.

The architecture is a three-way rental. The foundation model is Anthropic's Claude. Freshness comes from Microsoft's Bing grounding API. The differentiator is Yahoo's own knowledge graph: a billion-plus entities, 500 million user profiles, 18 trillion consumer events a year, and thirty years of search history. It cites sources transparently, so it can send traffic.

Amazon Rufus, which matters enormously or not at all

Rufus is the exception to everything above, because it does not answer from the open web. It answers from Amazon's catalog, listings, and reviews. No website work reaches it. If you do not sell physical products, skip it entirely; if you do, it may be the most consequential engine on this page.

Those figures are correlational — heavy Rufus users are plausibly heavy buyers already — but the gap is hard to wave away. The optimization surface is the listing: complete attributes, titles that name the specific use case, and review and Q&A content answering what a shopper asks right before buying.

Grok, DeepSeek, Meta AI, and the honest unknowns

Several engines have real usage and almost nothing verifiable behind them. Saying so plainly beats filling the gap with plausible-sounding guesses.

• Grok — 2.8% of US usage, third on First Page Sage's July 2026 list, and effectively undocumented. xAI publishes no crawler documentation, the tokens circulating in third-party bot trackers are unverified, and whether its web component uses Bing, Google, or its own crawl is not publicly known. It blends web results with real-time X data; past that, optimization advice for Grok is guesswork.
• DeepSeek — 0.4% of US usage, 0.02% of referral traffic. A real product with a negligible Western answer-surface footprint.
• Meta AI — Meta reported passing one billion monthly active users in Q1 2025, but that is distribution inside Facebook, Instagram, and WhatsApp, not search behavior. First Page Sage puts its US usage share at 0.05%, and its citation behavior and index source are both undocumented. Meta does publish crawler tokens, including Meta-WebIndexer, described as improving Meta AI search result quality.
• DuckDuckGo — DuckAssistBot crawls pages in real time for AI-assisted answers that, in DuckDuckGo's own words, prominently cite their sources, and the data is explicitly not used to train AI models. Opting out does not affect organic ranking.
• Mistral Vibe — publishes MistralAI-User and MistralAI-Index tokens and, notably, no training-crawler token. European relevance, minimal US share.
• Brave Leo — runs on Brave's own index, so it comes free with the Brave work you do for Claude.

The graveyard and the frozen

Half the engines in circulating roundups are not live targets.

• Phind shut down on January 16, 2026. It still appears in roundups published after that date.
• Arc Browser — active development halted in May 2025, and The Browser Company was acquired by Atlassian for $610 million in October 2025. Arc still runs and still gets security patches, but security patches are not a roadmap.
• Dia is a separate, AI-first browser from the same team, macOS and Apple Silicon only, with no committed Windows release. Size your expectations to that.
• You.com, Komo, and Andi still appear in vendor comparison tables. No current, verifiable usage data exists for any of them — an absence of evidence, which is a poor reason to spend a sprint.

Where the work actually goes

The prioritization falls out of the map. Four indexes, ordered by how much surface each hour buys you.

1. Google indexing. One job covers AI Overviews, AI Mode, Gemini, and Siri. Google's position is that a page must be indexed and eligible in Search, and nothing more.
2. Bing indexing. One job covers Copilot, Yahoo Scout, and much of ChatGPT's grounding. Verify in Bing Webmaster Tools, submit the sitemap, enable IndexNow, clear the crawl errors.
3. Brave presence. The only route to Claude, and the one index where robots.txt will not remove you — that takes a noindex meta tag plus a re-fetch request at search.brave.com/submit-url.
4. Your Amazon listing, if and only if you sell products. Rufus will never read your website.
5. Everything else is downstream. Perplexity, Leo, DuckAssist, and Vibe are either fed by those indexes or too small to plan around. Recheck the roster in six months.

One caveat: it is four indexes today. The Siri deal moved a major surface between columns in a single announcement, and Yahoo Scout added a 250-million-user surface by renting all three of its components. The map is stable enough to plan a quarter around. It is not stable enough to laminate.