QWE AI Academy The traffic drop that forced me to figure out AI Overviews
Last quarter one of our informational guides lost about a third of its clicks overnight on queries that started showing Google AI Overviews. Rankings barely moved. The overview sat on top, answered the question, and most people never scrolled to the blue links. That moment is why you should care about how to optimize for AI Overviews – zero-click risk is real, yet cited pages still get brand visibility and often higher-quality visits when someone does click.
I opened the official docs expecting a secret checklist. There wasn’t one.
Quick context: what Google actually requires
Eligibility is boring on purpose. Per Google’s AI features documentation, the page must be indexed and allowed a normal Search snippet. That’s the bar for showing up as a supporting link in AI Overviews or AI Mode. No extra technical membership card.
How the feature builds an answer is less mysterious than vendor blogs pretend. The same docs describe query fan-out (related sub-queries) and retrieval grounded in Search ranking systems. Overviews show only when the system thinks the summary adds something beyond classic results.
The official generative AI optimization guide is blunt: normal SEO still decides eligibility. Helpful non-commodity content. Crawlable site. Important facts in text, not locked in images or tab widgets. Structured data only when it matches what people see. Pages that load cleanly and work on mobile. llms.txt, AI-only markup packs, special machine-readable “AI files” – Google Search does not use those for these features.
Hands-on: rewriting one page for extractable passages
I stopped treating the underperforming guide like a ranking project and treated it like a retrieval target. Citation checks on our target queries looked better within a few weeks after recrawl – not because of a magic schema flag.
Mapped the head query plus three fan-out follow-ups people ask after they read a summary. Then rewrote each major section so the first 40-60 words could stand alone. No “as mentioned above.” Subject in the sentence. Claim + short reason. Detail and lists after.
// Before (buried)
Many factors go into good lawn care. Soil type matters...
// After (self-contained opener under H2 "How do I fix a weedy lawn?")
Fix a weedy lawn by identifying the weed type, applying the right herbicide or manual removal, then thickening turf with proper mowing and watering. Broadleaf weeds often need selective herbicide; grassy weeds need different chemistry. Test soil pH first if growth is patchy.That opener is citable without the rest of the article. Comparisons became simple HTML tables. Processes became ordered lists – both survive extraction cleaner than prose soup. I dropped in one original note from our own checks (“in three local soil samples the pH fix cut weed return by half”) so a generic model summary couldn’t fully replace the page.
Pro tip: Read each section aloud as if it were the only text the model sees. If it needs the previous paragraph to make sense, rewrite until it doesn’t.
Author byline stayed. Publish/update dates stayed visible. Internal links to deeper pieces stayed. No new schema beyond the FAQ markup that already mirrored visible Q&A. Search Console → request indexing → wait for recrawl.
Common pitfalls that waste weeks
- Thin pages for every long-tail fan-out variant. That pattern collides with scaled content abuse rules, and Google already expands multi-topic pages through fan-out without exact-match URLs for every sub-question.
- llms.txt worship and “AI schema” bundles. Official guidance: not required, not a lever for AI Overviews eligibility.
- Believing a top-10 organic rank still predicts citation. Ahrefs samples (as of the mid-2025 → early 2026 window) showed top-10 overlap for citations falling from roughly 76% to roughly 38%. Passages get pulled from deeper URLs and fan-out paths. Self-contained wording and clear entities beat position superstition.
- Answers trapped in tabs, client-only rendering, or image-only graphics with no text equivalent. If it isn’t textual and crawlable, don’t expect clean extraction.
Citation is not a full traffic refund. Plenty of people read the overview and leave. What you often keep is brand association plus the smaller set of higher-intent clicks.
What the numbers actually look like
Position 1 can lose half or more of its CTR when an AI Overview appears. Ahrefs and Seer Interactive analyses in the late 2025-mid 2026 period put informational drops around the high-50% to ~61% range depending on the study cut. Branded queries sometimes hold up better. Cited brands tend to lose less relative click share than brands that never appear in the overview links.
Google notes those clicks skew higher quality – longer engagement – but you still have to catch them in messy reporting. AI Overview / AI Mode activity rolls into Search Console’s Web performance views. Newer generative impression reporting (rolling out; some UI needs expansion) can show when your link was displayed. Clicks still look like normal organic. Perfect before/after isolation? Rare without query filters, annotations, and Analytics side-by-side. Watch branded search lift and engagement, not only raw clicks.
When you should skip heavy optimization
Navigational and straightforward transactional queries that almost never trigger overviews? Don’t burn a rewrite cycle. Already unique, structured, crawlable content that’s ranking? Stay fresh and leave it alone. Thin pure-commercial pages seldom feed complex informational overviews anyway. YMYL with weak expertise signals? Fix sourcing, authors, and evidence before you obsess over 50-word openers – structure won’t rescue untrustworthy copy.
FAQ
Do I need special schema or llms.txt to appear in AI Overviews?
No. Indexed + snippet-eligible is enough. Structured data only if it mirrors visible content.
Will ranking #1 guarantee my page gets cited?
No. Older samples showed tight top-10 overlap; later Ahrefs cuts near ~38% (early 2026 window) show fan-out pulling passages from a wider set. If you only chase position, you optimize for a correlation that already slipped. Open an incognito window, trigger the live overview for your query, and note which URLs actually get linked – that’s the ground truth, not your rank tracker alone.
How do I know if my changes worked?
GSC impressions on the URL (plus generative impression reporting if you see it). Manual checks across several days – overviews fluctuate. Branded search volume and time-on-page from organic sessions in Analytics. Example: we annotated the recrawl date, then compared the next two weeks of branded lift against a quiet prior window instead of trusting a single-day spike. Full surgical isolation of AI Overview clicks is still imperfect in the current Search Console setup, so use converging signals. If nothing budges after a real recrawl window, that query may not trigger overviews often enough to justify more labor.
Pick one underperforming informational page this week. Rewrite section openers into self-contained 40-60 word answers, add one original data point only you have, request indexing, then re-check live overviews after recrawl. Fastest feedback loop I know.