AI Overviews Aren’t Killing Your Traffic. They’re Re-Sorting Who Gets Cited.

The doom narrative around AI search is half-right and half-lazy. The half that’s right: informational organic traffic is in measurable, large decline, and the data has been triangulated by enough independent studies that arguing about it is no longer interesting. The half that’s lazy: framing this as an extinction event misses what’s actually happening underneath. AI Overviews, ChatGPT Search, Perplexity, and Google’s AI Mode aren’t deleting visibility — they’re re-sorting it. The passages getting retrieved into answers are coming from sites with citable, well-structured content, third-party validation, and a strong brand-search baseline as a proxy for entity coverage in the underlying corpora.

This piece is a field read for marketing leaders trying to separate the real shifts from the consultant theater. We’ve spent the past year working through the primary research, vendor studies, and practitioner POVs that matter — and watching what actually moves visibility for clients. The short version: most of the new “AI SEO” playbook is old SEO done seriously, plus a few genuinely new moves around answer formatting, schema discipline, community presence, and earned media. The longer version is below.

01 / The state of playAI answers are now the default surface — but not on every query.

Google AI Overviews moved from experiment to default across 2025. BrightEdge tracking shows AIO presence climbing from ~20% in early 2025 to roughly 48% by early 2026 — meaning roughly half of tracked queries still trigger no overview at all. Coverage skews heavily to informational queries and to Healthcare, Education, and Insurance verticals; e-commerce coverage actually contracted by 7.6 percentage points over the year, which suggests Google is still nervous about transactional intent and prefers to route shopping queries to clickable surfaces.

AI Mode is a different animal. Launched as a Labs experiment in March 2025, opened to all U.S. searchers at I/O in May, and pushed to 180 countries by fall. It returns a full-page synthesized narrative with linked citations and supports follow-ups — closer to ChatGPT than to a SERP. Sundar Pichai has indicated AI Mode is on a path to greater prominence inside Search; whether and when it becomes the default UI is still speculative. Worth tracking, not a planning input yet.

The other surfaces have all gotten bigger. ChatGPT remains OpenAI’s fastest-growing surface — total weekly active users (across all uses, not Search-only) crossed several hundred million through 2025, and OpenAI launched Shopping Research on November 24, 2025 as a product-comparison feature. Perplexity hit ~$200M ARR on subscriptions but quietly stepped back from its ad product in late 2025; reporting through early 2026 indicates it is leaning fully on subscription revenue (treat as reported, not a confirmed permanent exit until Perplexity says so on the record). Anthropic added web search to Claude in March 2025 and made it global on all plans by May 27. Microsoft Copilot is still the AI-search-first chat surface bolted onto Bing.

The click-through math is the part nobody disputes. Pew Research’s March 2025 study (n=68,879 searches, 900 U.S. adults) found that only 8% of users clicked a traditional result when an AIO was present, vs. 15% without; 26% of AIO sessions ended in zero clicks vs. 16% otherwise; less than 1% clicked links inside the AIO itself. Ahrefs’ updated December 2025 read of 300,000 keywords found AIOs cut position-1 CTR by 58% — up from the 34.5% they reported in April. Seer Interactive’s September 2025 update pegged the organic CTR drop at 61% (1.76% to 0.61%). SparkToro/Datos data put U.S. zero-click search at roughly 60–65% of all Google queries through 2025, with most of the remaining clicks going to branded and navigational terms. Google’s Liz Reid pushed back in August 2025, arguing that total click volume to the open web is “relatively stable,” that average click quality has slightly increased, and that the third-party studies rely on flawed methodologies, isolated examples, or pre-AIO traffic baselines. That’s a serious counterposition. The reason it doesn’t fully land: Google has not released aggregate data to support it, and DOJ filings have surfaced internal numbers consistent with the third-party reads.

The honest synthesis: top-of-funnel informational organic is in real decline; mid-funnel comparative and branded traffic is roughly flat to up for brands with strong equity; AI-referred traffic is small but high-converting where it exists (Adobe and Digital Commerce 360 both report gen-AI referral traffic converts ~31% better than other sources for the brands receiving it). The strategic posture follows from that math: preservation of share in a smaller but more qualified pool, not chase of pre-AIO volume.

02 / Who’s getting cited (and why)Reddit, Wikipedia, YouTube — and the brands they happen to mention.

The single most useful dataset in this space is Profound’s analysis of 680M citations from August 2024 through June 2025. Pair it with Semrush’s 230,000-prompt multi-platform study (over 100M citations across ChatGPT, Google AI Mode, and Perplexity) and a clear pattern emerges: each engine has a dominant editorial source, and that source is wildly different by platform. A single-channel content strategy will not generate visibility across all three.

One reconciliation note before the table. The Semrush 40.1% Reddit / 26.3% Wikipedia figure in the stat strip above is an aggregate across platforms and verticals from a 150K-citation sample — it’s the share of citations from those domains within that blended pool, not the share of all LLM citations any one source captures. The per-platform Profound numbers below run an order of magnitude lower because they’re slicing a single engine’s citation graph, where the long tail is much wider and includes LinkedIn, Medium, YouTube, and others that compress the visible share of any one source. Both reads are correct; they’re answering different questions.

A worthwhile mechanism note on the 16.7% figure, because it’s repeated across vendor decks and almost always glossed wrong. AIO doesn’t run a single “top-10 organic” lookup. It decomposes the user query into multiple sub-queries — Google calls this query fan-out — and retrieves separately for each, then a grounding model selects passages from the union. A page cited from position 47 for the headline term is almost certainly ranking top-10 for one of the decomposed sub-queries. The implication is not “rank doesn’t matter”; it’s “passage relevance to the fan-out matters as much as headline rank.”

A few more nuances worth weighting. ALM Corp’s industry tracking — a C-grade source, but one of the few looking at this specifically — reports citations from top-10 pages dropped from 76% to 38% as Google diversified sources into early 2026. That’s a steeper read than BrightEdge’s 32.3% → 54.5% organic overlap number (which is moving in the same direction but more slowly), and worth weighting as one practitioner-side data point rather than a settled finding. Reddit’s URLs ranking in Google jumped from 22M to 41.1M in the period following the Google licensing deal — a correlated but probably multi-causal shift, since Reddit’s own SEO revamp and Google’s broader forum-result push overlapped the same window. And industry tracking, including by Glenn Gabe, reported that a ChatGPT model update in October 2025 cut average brand mentions per response from roughly 6–7 to 3–4 — practitioner observation, not a published dataset, but a useful reminder that the citation surface itself is volatile.

03 / The brand visibility playbookOld fundamentals, sharpened. A few new moves that actually work.

The practitioners worth listening to in this space — Aleyda Solis, Lily Ray, Glenn Gabe, Eli Schwartz, Cyrus Shepard, Brodie Clark — are converging on roughly the same playbook. Vendor research from Frase, AirOps, Profound, Omniscient, and Semrush keeps confirming the same handful of moves. The work below is the connective tissue; the four findings most likely to surprise even an experienced SEO team are in the grid that follows.

Lead with answers, not setup — old discipline, newly load-bearing. BLUF (bottom-line-up-front) isn’t new. Featured-snippet optimization has recommended 30–60-word direct answers at the top of an H2 since at least 2017, and any content team that has shipped meaningful volume in the last five years already does this. What’s new is the cost of skipping it. Pre-AIO you lost a featured snippet; post-AIO the page often loses the citation entirely. Princeton’s GEO research (Aggarwal et al., ACM KDD 2024) found 44.2% of citations come from the first 30% of a page, and Stanford’s “Lost in the Middle” (Liu et al., TACL 2024) showed LLMs underweight content placed in mid-document. Frase and AirOps practitioner data — vendor-published, not peer-reviewed, and at odds with some of the schema correlation work below — both report FAQ-formatted pages with FAQPage schema get cited 40–60% more often than unstructured equivalents in their samples, and pages with three or more schema types show roughly 13% higher citation likelihood.

Be the obvious entity for your topic. Topic clustering and entity-based SEO are not new concepts; what’s new is that LLMs reward this work more visibly than ten blue links ever did. Comprehensive coverage of brand, product, ingredients, and use cases — across the brand’s own site and across third-party mentions — is what makes the brand the entity a retriever is most likely to surface a passage for. Aleyda Solis’s framing is right: technical implementation means nothing if the underlying strategy isn’t sound.

Schema is plausibly helpful, definitely not a lever. The evidence on schema is genuinely mixed. A December 2024 Search Atlas study found no correlation between schema coverage and citation rates. A September 2025 Search Engine Land head-to-head test (n=2 — directionally interesting, statistically near-meaningless) found the more-marked-up page won AIO inclusion. Growth Marshal’s analysis of 730 citations found attribute-rich schema cited 61.7% of the time, while minimally-populated schema has not outperformed having no schema in the limited samples studied (41.6% vs. 59.8%). Microsoft’s Fabrice Canel publicly confirmed at SMX Munich that schema helps Microsoft’s LLMs parse content. The reconciliation: treat schema as table stakes for parsing, not as a citation tactic. Implement it, populate it accurately, and don’t expect markup alone to move citation rates. For most DTC operators, the priority order is Product (with brand, GTIN, aggregateRating, review) first because it’s the schema with the clearest commercial use case in ChatGPT Shopping Research; then Organization (with sameAs to LinkedIn, Wikipedia, Wikidata); then FAQPage on top-trafficked informational pages; then Article with author and datePublished.

Earned media is the citation correlate. Omniscient Digital’s analysis of 23,000+ AI citations is the most-quoted dataset in this space because it produced three reinforcing findings from one study: brand search volume correlates with LLM citations at 0.334 (the strongest correlation in their sample, though weak-to-moderate in absolute terms — explaining roughly 11% of citation variance); 85% of brand mentions in LLM responses come from third-party pages, not the brand’s own domain; and 48% of citations on branded queries come from earned media. The same study also reported brands appearing in “best of” and comparison roundups are ~4x more likely to be recommended by LLMs than blog-only brands. One vendor analysis, treated as one data point — not four. The retrieval-mechanics read is the right way to frame it: branded search isn’t a lever you can pull on retrieval directly. It’s a downstream proxy for the upstream conditions retrievers actually see — entity coverage in training corpora, third-party editorial mentions, Wikipedia/Wikidata presence, and brand-and-category co-occurrence. Treat as directional until replicated. The operational implication still holds in trust-heavy verticals: a Wirecutter-tier roundup placement plausibly competes with a top-three organic ranking on the same query for citation share. In transactional categories with thinner editorial coverage, the calculus shifts back toward owned organic.

One honest caveat on the digital PR pitch. Tier-one editorial roundups (Wirecutter, NYT, Forbes-staff) are real, hard, slow, and not buyable — the realistic supply for any given category is a handful of placements per year across all brands competing. The accessible program is the much larger long tail of B2B-trade and niche-publisher roundups, where genuine editorial pickup is plausible if the brand has a real product story. A third category — listicle-mill brokers selling “guaranteed roundup placements” or paid inclusion in template comparison posts — is snake oil and should be treated as such. The signal LLMs respond to is editorial-grade pickup, not paid lookalikes of it.

One caveat on programmatic SEO. Eli Schwartz’s clickstream analysis of 260B+ sessions found Google and ChatGPT journeys run in parallel, not as cannibals — but he warns that AI content without strategy is not a shortcut. Programmatic pages should solve a real product or data problem (Netflix-style catalog, genuine comparison utility), not template-spam. The helpful-content systems will demote the latter, and LLMs will not surface passages from it.

04 / The measurement problemLift, not last-click. Honest tools, modest expectations.

AI-mediated traffic is genuinely hard to attribute, and pretending otherwise is how teams end up paying for tools that promise more precision than the surface allows. Five structural problems: the answer is often the destination — users get what they need in-chat from a synthesized answer with three citations and never click any of them, so there’s no traffic to misattribute because there is no traffic; chat answers that do route traffic often resolve in-app without a referrer; GA4 lumps unidentifiable AI sources under direct or “(not set)”; referrer strings from ChatGPT, Perplexity, and Claude are inconsistent and have changed multiple times in the past year; AIO impressions don’t reliably register in Google Search Console.

SparkToro’s Rand Fishkin has been arguing for years — and louder in 2025 — that lift-based measurement is more honest than per-channel attribution for this category. He’s right. Branded search volume, direct-traffic baseline shifts, and segmented geo holdouts give you a defensible answer when channel-level numbers don’t. We’ve used the same approach to size paid media incrementality for clients; it works for AI visibility for the same reason — and the in-chat-answer dynamic is structurally why per-channel attribution will keep falling short.

The tooling category is exploding. Profound (~$499/mo Lite, enterprise-tilted, deepest source-decomposition analytics), Peec AI (€89–499/mo, prompt-coverage with content recommendations), and Otterly.AI (~$29/mo entry, broad coverage of AIO/AI Mode/ChatGPT/Gemini/Copilot/Perplexity) are the established three; Promptwatch and Ahrefs Brand Radar are the most-watched competing entrants. Important caveat: SparkToro and Gumshoe.ai have both shown LLM responses are non-deterministic, so any “rank” reported by these tools should be read as distribution across N runs, not a position number.

A reasonable measurement stack for most marketing teams: GSC (track impressions-vs-clicks delta as a zero-click proxy), GA4 with segmented traffic-quality monitoring, branded-search volume in Glimpse or Google Trends, one AI-visibility tool sized to the budget, and a quarterly geo-holdout test for major content investments. That stack will tell you whether your visibility is moving. It will not give you per-prompt attribution, and you should be deeply suspicious of anyone selling that.

05 / What’s snake oilFour categories of marketing theater to ignore.

One more category to flag: “just write content for LLMs.” It’s thin advice, and dismissing it isn’t the same as dismissing the playbook. Roughly 80% of effective GEO is good SEO done seriously — but the remaining 20% is real and worth investing in. Specifically: BLUF-shaped answers at the top of every H2/H3 (an old discipline that’s gotten newly load-bearing), attribute-rich schema (treated as parsing table stakes, not a lever), a deliberate Reddit and community footprint executed with eyes-open about brand risk, and earned-media placement engineered for citation rather than for awareness alone. Those are the moves. Do them with the same rigor you’d apply to a technical SEO audit.

06 / What we’d watch for nextThe 2026 surface is going to keep moving.

Four developments worth tracking through the next few quarters. First, ChatGPT Shopping Research (launched November 24, 2025) is the most credible direct threat to Google’s transactional surface in a decade — watch how DTC brands with strong product-page schema and trusted comparison-roundup mentions perform inside it. Second, Perplexity’s reported step-back from its ad product (late 2025) means the company is leaning on subscriptions and Comet; downstream effects for brand visibility there are still developing. Third, Reddit’s licensing deals (Google and OpenAI, deal sizes reported in industry press) are correlated with Reddit’s outsized share of LLM citations — the causal share versus Reddit’s own SEO and Google’s broader forum-result push is genuinely contested, but the licensing-as-a-floor argument is plausible. Fourth, AI Mode’s continuing rollout is the one to model for: a future default-UI shift inside Google Search would reshape what “ranking” means. Worth a quarterly internal check-in, not a panic.

Bottom line below.