My Team Is Losing Traffic to AI Overviews and We Don't Know What Content to Fix First

Fix the pages that lost citations on money queries, not the ones that lost the most sessions. Informational guides absorb the biggest traffic drops when an AI Overview appears; those drops look dramatic in Analytics and often sit on pages that were never close to revenue. Comparison, alternative, and "best X for Y" URLs that lost the citation are the ones that change pipeline.

As of August 2026, that split is what the citation-mix data implies and what most teams get backwards.

Why the biggest traffic drop is the wrong queue

Promptwatch's July 2026 AI Overviews citation-type report is the better starting point for a fix queue. Listicles averaged about 18% of Overview citations in July (they were about 26% in Q1). Product pages averaged about 16.3%, and from July 28 they were the most-cited format each day (17.9% versus 16.2% for listicles). ChatGPT Search moved earlier: product pages were about 32.8% of citations in July, nearly double the ~18% share in March, while listicles sat at about 9.7%. Seer Interactive, through February 2026, still helps on leftover clicks: organic CTR on AIO-present queries recovered toward ~2.4%, and cited brands earned about 120% more clicks per impression than uncited brands. Pew's 2025 work: users click organic results on ~8% of AIO-present visits vs ~15% without.

Those numbers hit "what is X" content first. Overviews are densest on informational SERPs. A 10,000-session glossary page that loses 4,000 sessions can still be doing its job if it was never a conversion URL. A 800-session alternatives page that lost the Overview citation on a commercial prompt can be a quieter, more expensive problem.

Google's own AI features documentation still says there are no extra technical requirements beyond being indexed and snippet-eligible. The people-first guidance has not been replaced. You are not going to "SEO harder" your way out of a citation loss by rewriting the glossary first.

The score: commercial value × citation loss

Rank every affected URL with two factors. Multiply them. Work top-down.

Commercial value (1–5)

Pure definition / history / "what is"
Educational with a weak CTA
Category or use-case page
Comparison, alternatives, pricing, "best X"
Bottom-funnel: vs pages, implementation, "X for [job]"

Citation loss (0–3)

0: still cited in the Overview / AI answer on the target prompt
1: cited on some engines, missing on the one that sends you traffic
2: uncited, but a competitor or third-party page about you is cited
3: uncited, and a competitor's own page is the source

Priority = value × loss. A 5 × 3 page beats a 1 × 3 page every time, even if the 1 × 3 page lost 10× the sessions.

You cannot compute citation loss from session delta. Analytics tells you traffic moved. It does not tell you whether Google, ChatGPT, or Perplexity stopped naming you. Search Console's generative AI reports (shipped June 3, 2026) show impressions in AI Overviews, AI Mode, and Discover. The first version is impression-heavy; query and click coverage is limited. Google's help doc is the right place to read the caveats.

After GSC, you still need a prompt-level view. That is the job of a GEO tracker, not an SEO rank tracker. In the same week you pull GSC, put the money-query prompts into Promptwatch (or Profound / Otterly if you already pay enterprise prices) so citation loss is a number, not a vibe. Our tools list is the comparison; the method below is what you do with the output.

Informational pages get hit hardest — and often should wait

Treat informational traffic loss as expected compression, not an emergency rewrite list.

If the page still ranks, is still cited, and never converted, log it. Do not spend a sprint "adding more comprehensive coverage."
If the page is uncited and it is the only asset that could earn a citation on a head definition query that feeds your category, then it is worth an extractability pass: answer block, tighter headings, entity-consistent naming.
If the page is a thin affiliate roundup competing with Reddit and vendor docs, you will not win the Overview by publishing version 14. Earn mentions on the sources models already cite, or accept the CTR haircut.

Branded queries can move the opposite direction. When the Overview names you, leftover clicks concentrate. Seer's cited-vs-uncited split is the mechanism. Do not "fix" a branded URL that gained share because a non-branded URL next to it bled sessions.

Build the queue in one afternoon

Export the drop set. Last 90 days vs prior 90: URLs where clicks fell ≥25% and average position did not collapse. Tag intent.
Overlay generative impressions. GSC generative AI performance plus Bing AI Performance (Copilot / Bing AI citations, public preview since February 2026).
Map each money URL to 3–5 prompts a buyer would actually type or speak. Not your title tag. The query.
Score citation loss on those prompts across Google AI Overviews and at least one chatbot surface. Screenshot it. Arguments about "I thought we were in it" waste more time than the tracking.
Sort by value × loss. Cap the first sprint at 10 URLs. Extractability and source-earning, not a sitewide refresh.
Re-measure citations in 14 days, then clicks in 30. Clicks lag. If you only watch sessions, you will declare failure on a citation that already came back.

Step 4 is where teams stall. Doing it by hand does not scale past a few dozen prompts. A prompt portfolio overlaid with citation losses is the actual artifact — Promptwatch tracks ChatGPT, Perplexity, Gemini, Claude, Copilot, and Google AI Overviews daily, with screenshots and Slack/email alerts, from $29/mo. It does not write the page. Semrush's AI Toolkit and Ahrefs Brand Radar can show brand-level presence; they will not give you that prompt × URL queue unless you build it.

What to change on a high-priority page

Google's AI optimization guide is still the official floor: crawlable text, helpful content, structured data that matches what the user sees. GEO adds the extractability pass:

Lead with the answer in 40–80 words, then evidence.
Use the same product and entity names the rest of the web uses. Models reconcile strings.
Put comparisons in tables. Put caveats in the same block as the claim, so the model does not lift the claim alone.
Cite primary sources. Pages that quote Promptwatch data reports, Google, or vendor docs get used as sources more often than pages that paraphrase them into mush.
Allow OAI-SearchBot if ChatGPT Search matters. It is the citation crawler; GPTBot is training. robots.txt changes take ~24h. IPs: openai.com/searchbot.json.

Do not block Googlebot to "protect" content. Googlebot controls crawl for Search, including AI Overviews. The GSC generative AI control can exclude you from Overviews and AI Mode; that is an opt-out, not a ranking lever.

What not to do

Do not sort the backlog by "sessions lost" in GA4 and start at the top.
Do not delete informational content because CTR fell. The page may still be the citation source for a commercial answer.
Do not chase a universal "AI ranking factor" list. Nobody published one. How we rank tools is about software, not a secret GPT scoring spec.
Do not wait for GSC click data to catch up before you act on a citation loss you can already screenshot.

FAQ

Should we pause publishing until we know which pages to fix?

No. Pause rewrites of the drop set until they are scored. New pages that answer a prompt you are already uncited on can still be the right move. Do not ship another definition post to "make up" lost informational sessions.

Why would branded traffic go up while the blog dies?

The Overview is compressing the SERP. If the brand is named, remaining clickers self-select. Informational queries are where Overviews are most common, so the blog takes the listicle-to-product-page shift (and Pew's smaller leftover click) while branded holds. That is not a reason to starve the blog; it is a reason not to treat blog session loss as the priority list.

How long after a content fix should we expect the citation back?

Treat two weeks as the re-measure window for prompt trackers, 30 days for Search Console clicks. robots.txt / SearchBot changes: about 24 hours to propagate, longer to recrawl. Anyone promising overnight Overview inclusion is selling something else.

Can we just look at which URLs lost the most impressions in the generative AI report?

Impressions tell you Google showed you. They do not tell you commercial value or whether you lost the citation versus a competitor. Use the report as a filter, then score.

What to do this week

Pull the 90-day click-drop URLs where position held. Tag each 1–5 for commercial value.
For every 4 and 5, write the buyer prompts (not the URL slug).
Record citation status on those prompts. If you cannot do it in a spreadsheet without lying to yourself, use a daily tracker from the GEO tools list.
Multiply value × loss. Pick ten pages. Extractability pass only.
Set a 14-day citation check and a 30-day CTR check. Do not reopen the informational rewrite pile until the money-query citations move.