AI search is eating your organic traffic: what to do this quarter

Your rankings held because ranking was never the thing that changed. A growing share of buyer questions now get answered inside the AI response, and the source that gets cited is the one that gets the click. The job this quarter is to become a cited source for the five buyer questions that matter most, and that is measurable per question, per engine, in 90 days.

Why did traffic fall while rankings held?

The clicks did not vanish. They moved inside the answer.

When a buyer asks “what’s the best project management tool for a remote engineering team,” a growing number of AI surfaces answer in place, name two or three vendors, and close the session. The buyer got what they needed. Your page ranked. You got no visit. How sharp that effect is varies a lot by vertical and query mix, so distrust any single “AI cut traffic X%” number that does not name its study, date and vertical.

Here is the part most plans miss: only about 18% of ChatGPT conversations trigger a web search at all; the rest are answered from model weights (Search Engine Land, citing OpenAI data, June 2026). Most of “AI search” is not search in any crawlable sense. A plan that only fixes crawlability, sitemaps and structured data addresses roughly 18% of the surface; the other 82% is answered from what the model already knows, which means your job is to be in the corpus the model learned from, not just the index. The ranking system and the citation system are different systems, and they reward different things.

Which of your queries are actually at risk: branded, or problem-led?

This is the diagnostic question, and most teams get it wrong because they only check their own name.

A brand can sit at roughly 70% visibility on branded queries: ask an engine “what does [your company] do” and it answers accurately. That number feels reassuring, and it is not the number that matters. The number that matters is visibility on problem-led, cold queries: “how do I reduce churn in a B2B SaaS product,” “what tools help engineering teams ship faster.” These are the questions a buyer asks before they know your name, and in our own scan data (Collimer, July 2026) the same brand that scores about 70% branded routinely scores 0% on the cold queries while a rival owns that surface outright.

The fix starts with knowing which bucket you are in, and you can find out by hand today; the last section is that exercise. For the mechanics of why the gap exists, see why your AI visibility is lower than you think.

What do you stop doing this quarter?

Stop treating AI visibility as one move. “Optimize for AI” implies a single surface, and there is no single surface: only about 11% of cited domains overlap between ChatGPT and Perplexity (citation analysis, June 2026). ChatGPT skews toward Wikipedia, editorial outlets and established references; Perplexity skews toward Reddit and fresh community content. A tactic that earns citations in one engine can be irrelevant in the other.

In practice:

• Stop chasing a single pooled “AI visibility score.” An average across engines hides being strong in one and absent in another.
• Stop treating SEO and GEO as rivals. Your search budget is still working: rankings still drive sessions for the majority of queries that do not get answered in place. Keep most of the budget; add GEO work on top rather than swapping it in.
• Stop measuring this quarter in sessions. The goal is presence in answers, which converts on a different path and a longer lag than a session does.

What do you start: the five buyer questions and the page for each

Pick the five questions a buyer asks in the 30 days before they would sign with you. Not “what is [your category]” but the specific, evaluative ones: “how does [your category] handle [workflow],” “what does [your category] cost for a 20-person team,” “how long does it take to implement.”

For each question, build one page that answers it directly:

• State the answer before the context, in the first two sentences, with a number or a named thing in the first paragraph.
• Carry at least one concrete number, date or named comparison per section.
• Include a definition block or a numbered list an engine can extract without reading the whole page.
• Write for the buyer who does not know your name yet, not the one already in your trial.

This is not a volume play. Five pages done this way outperform fifty pages of topic coverage, because the engines reward extractability, not coverage.

How do you get corroborated where the engines actually look?

Your own pages are necessary and not sufficient. On evaluative queries, engines are roughly 6.5 times more likely to cite a third-party domain than a brand domain, and about 85% of brand mentions in AI answers come from third-party pages rather than the brand’s own site (our citation analysis of evaluative query sets, Q2 2026). The engines want independent confirmation that what you say about yourself is true.

The corroboration that moves the needle: community threads where practitioners name your tool in a specific context; third-party review summaries that describe your use case in the buyer’s language; and technical documentation, READMEs and integration guides that say plainly what your tool does.

One receipt from running exactly this play: a Series B devtools company shipped a documentation-and-community wave over six weeks, updated READMEs across its GitHub repos, integration partner pages, and detailed answers in three community threads, and its citation rate on problem-led queries rose from 1.25% to 5.83% across five engines, a 4.6-point lift on a base that had been effectively zero. No new product, no press. For the mechanism, see third-party sources win evaluative queries.

How will you know in 90 days whether it worked?

Measure per question, per engine, never a pooled score.

Your baseline is a table you build by hand in the next section: five questions, three engines, named or not, cited or not. Fifteen cells. In 90 days you run the identical table and look for movement in specific cells. The signals that the work is landing:

• Named in at least 2 of 3 engines on at least 3 of your 5 questions.
• Third-party pages that mention you are being cited on at least one evaluative query.
• The five pages you built are being crawled and, where an engine retrieves live, cited.

What you are not measuring this quarter is sessions recovered. Presence in answers drives branded search, direct visits and referrals, and none of those show up cleanly in a sessions report for 60 to 90 days after the citations start. Set that expectation with your CFO now, in writing, so week six is not a crisis meeting. For how to run the measurement without fooling yourself, see how to measure AI visibility properly.

The one thing to do today

Open ChatGPT, Perplexity and Claude. Ask each of the five questions a buyer asks right before they would pick you. Write down whether you are named and whether your site is cited. Fifteen cells, thirty minutes, done. That table is the baseline everything this quarter is measured against.

When you want the same check run systematically, run a free scan; it runs your buyer questions across five engines, repeats them to account for model variance, and puts an interval on the result so you know whether a change is signal or noise. It takes about 90 seconds, and it is the measurement step for the baseline you just built by hand.