TL;DR: Why SEO is becoming a black box in 2026
Google removed the num=100 parameter in September 2025, AI Overviews now cover the majority of informational queries, and nearly 60% of searches end without a click according to SparkToro (2024). SEO data is fragmented. The solution: track presence, not just positions.
Key points
- The removal of num=100 led to 87.7% of tracked sites losing visible keywords (Semrush, 2025).
- AI Overviews appear on ~30% of US queries and reduce organic CTR by an average of 34.5% (Ahrefs, 2025).
- ChatGPT referral traffic to websites increased 25-fold in one year (Similarweb, 2025).
- Focus on GEO (Generative Engine Optimization), server logs, and incremental measurement.
What is the SEO black box and why are we talking about lost signals?
The SEO black box refers to the progressive erosion of measurable signals: since 2011, "not provided" has masked organic keywords; in 2025, Google removed num=100, causing the number of tracked keywords to drop by 87.7% on sites monitored by Semrush (2025). Tracking is becoming an approximation.
The four stacked layers of opacity
The loss of signal is accumulating. First, the historical "not provided." Then GA4, which applies sampling beyond 10 million events according to Google Support. Next, Consent Mode v2, which models traffic refusing cookies. Finally, AI Overviews, which answer without a clicked link.
The result? A B2B CMO's dashboard in 2026 no longer shows reality. It shows a silhouette, and each layer adds blur. SEO is no longer flying by sight: it's flying with degraded radar.
The "dark traffic" of LLMs
ChatGPT, Claude, Perplexity, and Gemini cite sources without always generating a click. When a click occurs, the referrer often appears as "direct." Similarweb (2025) measured a 25-fold increase in referral traffic from ChatGPT in one year, but attribution remains unclear on the GA4 side.
Why is data fragmentation accelerating in 2026?
Three forces are converging: AI Search is reshaping the SERP, privacy regulations are drying up cookies, and publishers are locking their content behind paywalls or aggressive robots.txt files. Cloudflare (2024) reports that nearly 40% of the top 10,000 sites now block GPTBot or Common Crawl.
AI Search redefines the SERP
AI Overviews appear on approximately 30% of US informational queries according to Ahrefs (2025). The average CTR for position 1 drops by 34.5% when an AI Overview is present. On Bing, Microsoft has integrated Copilot directly into the SERP: the barrier between search and generation is blurring.
Apple Privacy and the end of third-party cookies
Mail Privacy Protection, ITP 2.3, App Tracking Transparency: Apple has dismantled behavioral tracking on iOS. According to Flurry Analytics, only 25% of iOS users opt-in to app tracking. On the web, Google's Consent Mode v2 now models what is no longer measured.
Paywalls and bot blocking
The New York Times, Reddit (Google deal worth $60M/year according to Reuters, 2024), Stack Overflow: training data is being monetized. Sites that block AI crawlers disappear from the LLM's reference corpus. A page invisible to Claude will never be cited by Claude.
How to push the limits with server-side tracking?
Server-side tracking recovers 30 to 50% of the signal lost on the client side according to Simo Ahava, a reference on Google Tag Manager. The logic: send hits from your server (or a Cloud Run endpoint) rather than from the browser, bypassing adblockers and ITP.
Recommended server-side stack for 2026
A viable setup combines GTM Server-Side on Cloud Run, a first-party domain endpoint (analytics.yourdomain.com), and event duplication to a warehouse (BigQuery, ClickHouse). This architecture preserves attribution despite Safari ITP purging first-party JavaScript cookies after 7 days.
Server logs: the forgotten raw data
Log analysis (Nginx, Cloudflare, Apache) reveals exactly which URLs Googlebot, Bingbot, GPTBot, and ClaudeBot crawl. Search Engine Journal (2024) reminds us that an average site sees 50 to 70% of its crawl budget wasted on non-strategic URLs. Logs don't lie, unlike third-party tools.
How to track your visibility in LLMs (GEO and AEO)?
GEO (Generative Engine Optimization) consists of measuring and optimizing a brand's presence in ChatGPT, Claude, Perplexity, and Google AI Overviews responses. According to Profound (2025), brands cited in LLMs gain an average of 3.2x more share of voice than in the classic SERP.
LLM citation tracking tools
Profound, Otterly.ai, AthenaHQ, and Peec.ai monitor brand mentions on recurring prompts. You define a panel of 100 to 500 B2B prompts ("best growth marketing tool France," "AI SEO agency"), the tool queries them daily, and returns a share of voice per generative engine.
Optimize for citation, not click
LLMs cite what they understand. An arXiv study (Aggarwal et al., 2024) on GEO shows that content with sourced citations and numerical statistics gains 30 to 40% more visibility in generative responses. Structure your content for extraction: short sentences, named data, named sources.
Our on-page SEO methodology now integrates these LLM extraction patterns by default.
What alternative KPIs to adopt when positions are no longer enough?
When average position becomes noisy, refocus on robust indicators: branded query volume, direct traffic, LLM share of voice, earned mentions. Search Engine Land (2024) notes that branded queries have become the most reliable proxy for a B2B brand's SEO health.
The quartet of robust KPIs
Four indicators resist the black box: branded query growth (Search Console, filter by brand name), direct traffic (a proxy for notoriety), LLM share of voice (Profound or equivalent), and the number of earned mentions on tier-1 domains.
Incremental modeling and MMM
Open-source Marketing Mix Modeling (Meta's Robyn, Google's LightweightMMM) allows measuring SEO incrementality without cookies. You correlate investment, SEO signals (organic sessions, brand search), and conversions, and isolate the real contribution. It's slow, but it resists signal loss.
Comparative table of old vs. new school KPIs
| Old KPI | 2026 Limit | Replacement KPI |
|---|---|---|
| Average position | Num=100 bias | LLM share of voice |
| Top 10 keywords | Partial sampling | Branded search growth |
| Organic traffic GA4 | Sampling + dark traffic | Direct traffic + server logs |
| SERP CTR | Captured by AI Overviews | Tier-1 earned citations |
How to combine digital PR and SEO to gain entity authority?
LLMs rely on named entities and their co-occurrences. A brand mentioned 50 times on tier-1 sources is 4 to 6 times more likely to be cited by ChatGPT than a brand without press according to Profound (2025). Digital PR is no longer a complement: it's the fuel for GEO.
Source what goes into LLM corpuses
Common Crawl, the main dataset feeding GPT and Claude, massively indexes Wikipedia, Reddit, Stack Overflow, .gov, .edu sites, and established press. A mention in Les Echos or TechCrunch weighs infinitely more than a guest post on an unknown blog. Aim for authority quality, not link quantity.
Reinforced E-E-A-T triangle
Identified author with author page, schema.org Person, consistent LinkedIn profile, external mentions: this is what constitutes an entity recognized by algorithms. Google Search Central has insisted since 2022 on demonstrable lived experience. LLMs apply similar logic: they cite sources they can attribute.
To structure this authority, see our entity-oriented SEO agency approach.
What concrete 90-day action plan to regain control?
Within 90 days, a minimal viable setup can be deployed. According to Gartner (2024), 64% of CMOs are now investing in alternative measurement tools. Late adopters lose visibility without even knowing why.
Day 0-30: Audit and baseline
Audit your Search Console post-num=100 (annotate the September 2025 break), export 12 months of server logs, list 200 relevant LLM tracking prompts for your business, and benchmark your competitors on Profound or Otterly. Identify tier-1 citation gaps.
Day 30-60: Technical deployment
Activate GTM server-side on a first-party subdomain. Connect your server logs to a parser (Screaming Frog Log Analyzer or Botify). Launch weekly LLM share of voice tracking. Document your AI Overviews screencaps by strategic query.
Day 60-90: New reporting
Redesign the CMO dashboard around four blocks: brand health (branded queries, direct traffic), generative presence (LLM share of voice, captured citations), authority (tier-1 earned mentions, contextual backlinks), incrementality (MMM model if sufficient volume). Communicate the methodology to stakeholders.
To equip this plan, our artificial intelligence agency automates the collection of LLM citations via API.
FAQ
What does "black box" mean in SEO?
The term "black box SEO" refers to the increasing opacity of SEO data: removal of num=100, AI Overviews capturing clicks, GA4 sampling, LLM dark traffic. Nearly 60% of Google searches now end without a click according to SparkToro (2024).
How to measure your visibility in ChatGPT or Claude?
Tools like Profound, Otterly.ai, AthenaHQ, and Peec.ai daily monitor your target prompts on ChatGPT, Claude, Perplexity, and Gemini. You get a share of voice per engine. Profound (2025) indicates that tracked brands detect an average of 40% more mentions invisible in the classic SERP.
Is classic SEO dead in 2026?
No. Google still processes approximately 5 trillion queries per year according to Google (2025). The organic channel remains the primary driver of B2B traffic. But its traditional KPIs (positions, CTR) are eroding. SEO is evolving into a discipline of entity authority and multi-surface presence.
Should GPTBot and ClaudeBot be blocked in robots.txt?
It all depends on the business model. Cloudflare (2024) notes that 40% of the top 10,000 block these bots. If your revenue comes from clicks and ad-tech, blocking makes sense. If brand visibility in LLMs matters, allow them: a blocked site will never be cited.
Which tools should be prioritized for a black box-proof setup?
Minimal viable stack: GTM server-side, BigQuery, Screaming Frog Log Analyzer, Search Console API, an LLM tracking tool (Profound or Otterly), and ideally an open-source MMM like Meta's Robyn. Expect €1,500 to €5,000 per month depending on volume.
