State of AI Search — July 2026

rocketblue’s monthly read on how AI engines actually cite, mention, and rank brands. Aggregate data across every brand and organization monitored on the rocketblue platform — not one company’s numbers, but the market’s.

The 60-second version

• ChatGPT cites a source in 94.7% of responses; Claude in just 17.4% — different engines, different disciplines, not one “AI visibility” score.
• Google’s AI Overview citation mix shifted hard this month: gstatic.com dropped from 9,402 citations to zero, while vertical-authority sites (legal directories, gov sites, research repositories) posted their best month in our data.
• The median AI-cited page runs 1,428 characters — short and structured, not a 3,000-word guide.
• 71.4% of brand mentions in AI answers were positive. The real risk isn’t bad press — it’s the 24.5% of responses that named no brand at all.
• Only 18.96% of citations show up inline in the visible answer. Most are footnotes nobody reads.

The headline

Google’s AI Overview quietly changed how it cites images this month. gstatic.com — the domain behind Google’s static image assets — went from 9,402 citations in June to zero in July. At the same time, vertical authority sites had their best month in our data: attorneys.superlawyers.com (+2,263), techradar.com (+2,235), gov.uk (+1,772), arxiv.org (+1,619), g2.com (+1,284).

We can’t say for certain these two things are the same story. But the direction is consistent: less generic infrastructure, more specific expertise. If you publish content that reads like a credentialed answer instead of a web page, this is the kind of month that rewards you.

The citation landscape

YouTube (39,422 citations), Reddit (31,516), and Google (22,928) remain the three most-cited domains across all engines this month — but all three declined from June. Reddit alone dropped 20,429 citations. YouTube dropped 12,373.

That doesn’t mean UGC is dead. It means the composition of what gets cited is shifting month to month, which is exactly why this is a monthly report and not a static claim. Domains that gained the most ground were almost all narrow and credentialed: legal directories, gov sites, review platforms, research repositories. Nobody’s homepage-generalist site is on that gainers list.

Takeaway: if your content strategy depends on being the biggest, general-audience resource in your category, this month’s data says that’s a shrinking lane. Specific, credentialed, narrow beats broad.

How the engines actually differ

This is the part most GEO advice skips: the six major AI engines don’t cite the same way, at all.

ChatGPT cites almost every answer, from the widest pool of domains of any engine. Claude is the opposite: it cites fewer than one in five responses, and when it does, it pulls from a strikingly narrow set — just 3,208 domains all month, dominated by a single vertical (B2B finance/AP automation software: highradius.com, cashbook.com, serrala.com, medius.com). Optimizing for “AI visibility” as one undifferentiated goal doesn’t hold up against numbers like these. Getting cited by ChatGPT and getting cited by Claude are close to different disciplines.

Brands: named, not necessarily won

• 24.5% of all analyzed responses named zero brands.
• 46.9% named 4-9 brands; 14.7% named 10 or more.
• Perplexity was the exception — median 2 brands per response, the least crowded field of any engine.
• Sentiment when brands were named ran heavily positive: 71.4% positive, 28.3% neutral, 0.3% negative.

The practical read: the risk in AI search isn’t reputational. It’s invisibility. Almost nobody gets mentioned and then trashed. They just don’t get mentioned.

What actually gets cited

We analyzed the properties of the source pages AI engines cited. A few things stand out:

• Homepages: 20.1% of citations — the single most-cited content type.
• Blog posts: 19.75% — combined with homepages, that’s more than product pages, guides, listicles, and comparison pages put together.
• Median cited page length: 1,428 characters — well under what most SEO playbooks recommend for “comprehensive” content.
• 83.2% of cited pages used bullet points.
• 66.6% included visual data.
• Only 8.2% presented original research.

One more distinction worth knowing: only 18.96% of all citations appeared inline in the visible answer text. The rest were reference-only — footnotes a user would have to click to ever see. Being cited and being seen are not the same metric, and most citation counts (including headline ones like the domain rankings above) don’t distinguish between the two.

(The median page length and structural figures — bullet points, visual data, original research — are based on a deep-dive sample of 368 analyzed source pages, not the full citation set: directional, not exhaustive. The homepage/blog citation-share figures are from the full dataset of 1.8M+ citations.)

What this means for your strategy this month

Short version: narrow beats broad, structure beats length, and which engine you’re optimizing for changes the playbook. If you’re chasing one undifferentiated “AI visibility” score, you’re already behind — the data says the engines don’t behave alike, and neither should your content plan.

We’ll be back with the August numbers in early September.

Frequently asked questions

Which AI engine cites sources most often? ChatGPT, citing at least one source in 94.7% of its responses — the highest rate of any engine rocketblue tracks. Claude has the lowest, at 17.4%.

What’s the most-cited domain across AI engines right now? YouTube, with 39,422 citations in July 2026, followed by Reddit (31,516) and Google.com (22,928) — though all three declined from the previous month.

How long should content be to get cited by AI? Based on rocketblue’s analysis of cited source pages, the median length is 1,428 characters — short and structured, not exhaustive. Length matters less than clear structure: bullet points, visual data, and a stated point of view correlate more strongly with citation than word count.

Does getting mentioned by AI risk negative brand sentiment? Rarely. Across all brand mentions analyzed in July, 71.4% were positive and only 0.3% were negative. The bigger risk is not being mentioned at all — 24.5% of analyzed responses named no brand.

Methodology: rocketblue sends a weekly set of prompts — system-generated and user-defined — to ChatGPT, Google Gemini, Google AI Overview, Google AI Mode, Perplexity, and Claude, replicating how a real user would ask a question in a given target country. Responses are analyzed for brand mentions, sentiment, and cited sources. Copilot and Grok are excluded — too little query volume on the platform to be statistically reliable. Response-volume figures reflect rocketblue’s own prompt-scheduling mix across customer brands, not the engines’ real-world usage share, and should not be read as market share. “Citations” refers to sources an AI response referenced, not clicks or traffic. All figures are final month-end numbers for July 1–August 1, 2026.