If your revenue depends on a top-three ranking for a category keyword, one number from BrightEdge's AI Overview tracking should change how you plan next year. In mid-2024, 76 percent of the pages cited inside Google's AI summaries ranked in the organic top ten. By February 2026, that share had fallen to 17 percent. The ranking you bought stopped being a ticket into the answer.

That shift explains a lot of the odd reporting coming out of e-commerce teams this year. Rankings hold or improve, impressions climb, and sessions fall. The dashboard is not broken. The page that used to end a search now feeds a model that answers the question on the results page, and the user never arrives.

What changed, in plain terms

Traditional search had one job: return a list of links ordered by likely relevance. You optimized for the order, and the order determined traffic. Generative search retrieves a set of candidate passages, weighs which ones are worth quoting, and writes an answer that cites some of them. Sometimes it cites three sources. Sometimes it states a fact and cites nothing.

There are now two competitions running on the same keyword. The first is for position. The second is for the citation, and it scores differently. Content that ranks eighth but states a clean, sourced, specific claim can get pulled into the summary while the page in position one gets skipped. The Princeton, Georgia Tech and IIT Delhi group that formalized this in their 2024 KDD paper on generative engine optimization found that a source ranked fifth could gain 115 percent visibility by citing its own sources, while the top-ranked page lost 30 percent. Position is an input to that contest, not the outcome of it.

The two numbers that look like they contradict each other

Organic click-through rates on queries with an AI Overview fell from 1.76 percent to 0.61 percent between June 2024 and September 2025 in Seer Interactive's longitudinal study, which tracked 2.43 billion impressions across 53 brands. First Page Sage measured a 58.6 percent drop in organic clicks per thousand impressions across its client set, with zero-click rates on those queries rising from 62.7 to 82.4 percent.

At the same time, Adobe Analytics reports that AI-referred traffic to US retail sites converted 42 percent better than non-AI traffic in March 2026. Twelve months earlier, that same channel converted 38 percent worse. AI visitors also produced 37 percent more revenue per visit and spent 48 percent longer on site.

Both numbers are true, and the reconciliation matters more than either one. AI Overviews absorb the top of the funnel, the browsing and comparing phase. What reaches your site from an assistant is a person who already did that work in a chat window and arrived with a shortlist. Fewer sessions arrive, and the ones that do are worth more. An analytics dashboard built to count sessions will read this as pure loss when it is partly a change in what a session means.

E-commerce got off lighter than publishing, and there is a reason

Media and publishing took the steepest damage in First Page Sage's industry breakdown, at 47.3 percent. An AI summary can fully satisfy "what year did this happen" without a click. Local services, real estate and retail lost far less, at 6.8, 9.2 and 11.6 percent. Buying something still requires a checkout, and a checkout requires a browser tab.

The risk for an online store is not that AI answers replace the purchase. It is that the model decides which three products make the shortlist before your site is ever loaded. That gate is narrower than page one of Google, and it is controlled by whoever writes the answer.

What the research says earns the citation

The GEO paper tested nine optimization methods across ten thousand queries and found the biggest lift came from adding citations, quotations and statistics: roughly 40 percent improvement in visibility. A model assembling an answer wants sentences it can lift without hedging. "Our solution is fast and reliable" gives it nothing. "Median checkout latency fell from 1.9 seconds to 640 milliseconds after we moved rendering to the edge, measured across 4.2 million sessions" gives it a sentence it can use, with a number attached.

Ahrefs' study of 75,000 brands found something that unsettled a lot of link builders. Brand mentions correlated with AI visibility at 0.664. Backlinks correlated at 0.218. Mentions win by roughly three to one. Language models read text, not link graphs. A brand discussed in an industry publication, a comparison thread or an analyst note becomes part of the model's picture of a category, and domain authority does not substitute for being named.

A product page, before and after

Take a category page that ranks fourth for "industrial label printer." The current copy opens with "Our industrial label printers deliver reliability and performance for demanding environments." That sentence could describe any printer built in the last twenty years, and a model has no reason to quote it.

The rewrite answers the question the buyer is actually asking, within the first hundred words, with specifics: which substrates the unit handles, at what speed, what the failure mode looks like at the rated duty cycle, and what it costs to run per label. Same page, same design, longer copy. The change is that every claim carries a number a model can lift and a human can check.

That is the whole technique at the page level. It is unglamorous, and it works because the model's job gets easier.

The technical prerequisites most sites fail

None of the content work matters if the retrieval crawler cannot read the page. Most AI crawlers do not execute JavaScript. If your product grid, specifications, pricing and reviews are assembled client-side after the HTML arrives, the crawler sees a shell. Googlebot renders JavaScript and will eventually index it. GPTBot, ClaudeBot, PerplexityBot and OAI-SearchBot mostly will not.

The fix is server-side rendering or static generation for anything you want cited: product pages, category pages, comparison content, specification tables. Check it the crude way. Fetch a product URL with curl, with no JavaScript, and read what comes back. If the response has no price, no SKU and no specification text, that is your citation ceiling.

Then check robots.txt, because the defaults work against a lot of stores. Crawlers split into two groups that people conflate. Training crawlers such as GPTBot, ClaudeBot, CCBot and Google-Extended collect text for model training. Retrieval crawlers such as OAI-SearchBot, Claude-SearchBot and PerplexityBot fetch pages to answer a live question and to build the index behind those answers. Blocking the first group is a defensible editorial position. Blocking the second removes you from the answer entirely. One survey of crawler configurations found that roughly half the sites blocking GPTBot also blocked OAI-SearchBot, usually by accident through a shared disallow list. Each bot needs its own user-agent group.

The retrieval bots that need an explicit allow, as of October 2026:

• OAI-SearchBot feeds ChatGPT search and its citations
• Claude-SearchBot handles Anthropic's search index, separate from ClaudeBot training
• PerplexityBot builds the Perplexity answer index
• ChatGPT-User and Perplexity-User fetch a single page when a person asks for it

A wildcard disallow does not override a specific bot's group, so the safe pattern is one block per user agent with an explicit allow, and the wildcard group reserved for admin, cart and account paths.

What does not work

Several things being sold as AI search optimization have no evidence behind them. The llms.txt proposal, a markdown file at the root that summarizes a site for models, has no confirmed adoption by any major provider. Pasting "AI-optimized" keyword blocks into product descriptions does nothing measurable. Schema markup helps when it accurately describes the page and hurts when it describes a page you wish you had.

The gap between what vendors claim and what retrieval systems actually do is wide, and the honest position is that the field is about eighteen months old. Nobody has a complete map of it.

A plan that fits in two quarters

Start with a citation baseline. Pick twenty queries your buyers actually type, run them through Google with AI Overviews, ChatGPT and Perplexity, and record who gets cited and what the answer says about your category. Repeat monthly. The work is manual, it takes an afternoon, and it is the only honest measurement available before the tooling matures.

Rewrite the pages that should be cited so each one answers a specific question in its opening paragraph, with a number or a source attached. Product pages need specifications in the HTML rather than in a tab that renders on click. Comparison and buying-guide content belongs as text on your own domain, not only inside a gated PDF.

The work, in priority order

Fix rendering and robots.txt before writing anything new, because an unreadable page cannot be optimized. Then work the content.

• Confirm the product and category templates return price, SKU and specification text in the raw HTML
• Give every retrieval crawler its own allow block in robots.txt
• Rewrite the twenty highest-value pages so the opening paragraph answers one concrete question with a sourced number
• Build the citation baseline and re-run it monthly

Then instrument AI referrals as their own channel. Filter by referrer hostname for chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai, and watch conversion rate separately from volume. Be aware that the referrer filter catches only part of the traffic. In Ethercycle's ten-store analysis, sessions carrying ChatGPT's own tag outnumbered sessions carrying a chatgpt.com referrer by three to one, because app traffic sends no referrer and lands in analytics as Direct. Tag the links you control and accept that the untagged remainder will always be an estimate.

Expect the volume to look trivial next to paid search for a while. That is normal for a channel in its second year, and it does not tell you much about where the channel will be in its fourth.

The honest caveat

Ethercycle puts AI-attributed revenue at roughly 0.12 percent of 2026 revenue across the stores it examined, maybe 0.19 percent if you credit the app clicks that hide in Direct. That is not a channel anyone can build a quarter on, and anyone selling AI search optimization as an immediate revenue replacement is selling you something else. The case for the work is structural rather than urgent. Discovery is migrating into AI answers at a rate that has surprised everyone measuring it, with Adobe reporting AI referral traffic to retail up 1,219 percent from October 2024. The sites that are readable and specific when that volume arrives will already be in the answer. The ones that blocked the retrieval crawler in 2025 will be arguing with their agency about why.

Rankings still matter. They are no longer sufficient, and the distance between ranking and being cited is where the next two years of search work happens.