Summary: This case study examines a realistic scenario where mid‑market B2B marketing teams see organic sessions fall while Google Search Console (GSC) reports stable rankings. Competitors are appearing in AI Overviews and conversational AI outputs (ChatGPT, Claude, Perplexity) while the brand is invisible there. Marketing leadership demands better attribution and ROI proof. We walk through background, the challenge, the approach, the implementation, measured results, lessons learned, and how you can apply them. The analysis is data‑driven and includes advanced diagnostic techniques and thought experiments to isolate causes and fixes.
1. Background and context
Audience: mid‑market B2B marketing teams (the placeholder ). Company profile: SaaS product serving finance teams, ARR $10–50M, steady SEO investment, content calendar producing weekly long‑form posts and feature pages. Tools in use: Google Search Console, Google Analytics 4, Ahrefs, internal CRM (Salesforce), and basic server logs. No formal program to track AI channel impressions or conversational AI presence.
Key business need: demonstrate attribution and ROI to justify continued SEO/content spend amid board pressure to prioritize paid channels. Recent observation: organic sessions down 23% year‑over‑year, while GSC shows median positions for target keywords unchanged (+/‑ 1–2 positions). Marketing asks: “Why are we getting fewer visitors if rankings are stable?”
2. The challenge faced
Symptoms and constraints:
- Organic sessions down 23% YoY; non‑brand organic traffic down 30%. Brand queries stable or up slightly. GSC shows stable impressions and median positions for target pages. Click‑through rates (CTR) in GSC have declined for several non‑brand queries. Competitors show up in Google’s AI Overview (the “AI‑powered snapshot” or “About these results” / “AI Overviews” depending on UI), while the brand doesn’t. No visibility into what large LLMs (ChatGPT, Claude, Perplexity) surface about the brand; marketing can’t audit or optimize for that channel. Marketing budget is being scrutinized; leadership wants clear attribution (MQL per channel, ROI per content piece).
Primary hypothesis: Search engine behavior and external AI summarizers are siphoning click volume away from organic SERPs (via enhanced SERP features, AI Overviews, or LLM answers), while GSC ranking metrics don’t capture those extractive interactions. The brand’s content is not being surfaced in AI summaries because of signals, entity recognition, and citation practices favoring competitors.
3. Approach taken
High‑level approach — three parallel tracks:
Diagnostic: establish precisely where clicks were lost — which queries, which topical clusters, and which SERP features replaced clicks. Competitive & AI presence audit: determine why competitors appear in AI Overviews and conversational LLM answers; measure brand presence in LLM outputs. Attribution & experiment design: build testable attribution models and run controlled lift tests to tie content to pipeline outcomes.Advanced techniques used:
- Query‑level CTR decay analysis in GSC linked to session logs — align impressions to clicks at a daily cadence and detect segments with CTR decline despite stable impressions. SERP feature change detection via synthetic queries and SERP snapshots over time (using a headless browser + Googlebot‑like user agent) to capture AI Overviews, People Also Ask (PAA), knowledge panels, and featured snippets. LLM probing framework — standardized prompt set sent to ChatGPT, Claude, and Perplexity to evaluate what answers are returned for brand‑relevant queries, and whether competitors are cited. Cohort attribution with geo or time holdouts (lift testing) to measure causal impact of content changes on MQLs and SQLs.
4. Implementation process
Diagnostic pipeline
Steps:
Exported GSC data at the query/page/day level for 12 months. Joined with GA4 session data mapped by landing page and day. Calculated delta metrics: impressions delta, clicks delta, CTR delta, sessions delta. Flagged queries and pages where impressions were stable but clicks and sessions dropped >15%. Segmented queries into brand vs non‑brand, and into intent buckets (informational, commercial, navigational) using patterns and page taxonomy.Key finding: Most of the traffic loss concentrated in mid‑funnel informational queries (e.g., “how to automate invoice approvals”), where impressions were stable but CTR fell by 28% and sessions by 34%.
SERP and AI audit
Procedure:
Created a list of 200 priority queries (top traffic drivers and high intent mid‑funnel queries). Ran daily headless browser snapshots for those queries across US results for 90 days to capture any new SERP features (AI Overviews, PAA, featured snippets). Logged each SERP feature presence and captured the top text of AI Overviews and snippet citations.Findings: For 42% of priority queries, Google began showing an AI Overview / “AI‑powered summary” that condensed answers and cited 1–3 sources — in 67% of those cases a competitor site (usually a large, topical resource or research landing page) was cited; our pages were not cited. When an AI Overview appeared, organic CTR for the top 5 results dropped 30% on average.
LLM probing
Method:
Standardized prompt templates (e.g., “Explain how X solves Y. Include product recommendations.”) run through ChatGPT, Claude, and Perplexity for the 200 queries. Automated parsing of responses to extract brand mentions, citation links (if any), and whether the LLM recommended competitors.Findings: LLMs frequently returned summaries citing public knowledge sources (industry reports, large editorial sites). Our brand was rarely suggested unless the prompt explicitly mentioned us. Perplexity surfaced competitor links in 48% of tests; ChatGPT and Claude provided factual summaries without citations by default but favored well‑known resources when asked for sources.
Attribution experiments
Design:
Implemented server‑side tagging to capture original referral and landing page experiment variants. Ran a geo holdout experiment: in a control region, left content and CTA unchanged; in the test region, we added structured data, explicit “recommended citation” blocks, short answer snippets in page HTML, and an “AI‑friendly summary” meta block designed for LLM extraction. Measured MQLs, sessions, and content‑to‑pipeline conversions over 8 weeks.5. Results and metrics
Composite outcomes after 8–12 weeks:
Metric Control Region Test Region (AI‑optimized + schema) Lift Non‑brand organic sessions -31% YoY -12% YoY +19 pp improvement vs control CTR on priority queries -28% -8% +20 pp MQLs attributed to content -22% -2% +20 pp Share of LLM citations (measured via probing) 3% 18% +15 pp Revenue influenced (weighted pipeline) -15% -3% +12 ppInterpretation: Targeted AI‑centric page modifications and schema signals reduced traffic decline and improved both CTR and downstream pipeline metrics relative to control. The largest effect came from regaining visibility in AI Overviews and being cited by LLM probes — not because rankings changed, but because extractive AI features began including the brand as a source.
6. Lessons learned
1) Rankings ≠ clicks when extractive SERP features or LLMs intervene. GSC’s position metrics measure where your URL appears, but not whether an AI summary or featured snippet is replacing clicks. Always pair rankings with CTR and session analysis.
https://faii.ai/insights/ai-seo-optimization-services-2/2) AI Overviews and LLMs favor content that is (a) well‑structured, (b) clearly authoritative on specific entities/topics, and (c) easily parsable for short summaries and citations. Large, comprehensive resource pages and studies often get cited.
3) You can optimize for AI presence without gaming. Practical signals include:
- Concise “summary” sections at the top of pages (50–150 words) labeled clearly (e.g., “Quick answer”). Structured data (FAQ, HowTo, Article, Dataset) and explicit attribution blocks (e.g., “Source: [Study], link”). Clear author and publication metadata and canonical references to primary data.
4) Attribution needs experimental evidence. Probabilistic models alone won’t satisfy skeptical finance stakeholders. Geo or time holdouts and server‑side event capture provide stronger causal evidence of content ROI.
5) LLM auditing is feasible: build a repeatable probing framework and record responses. This provides visibility into what ChatGPT/Claude/Perplexity are saying about your brand and where competitors are recommended.
7. How to apply these lessons
Step‑by‑step playbook you can replicate:
Run a GSC + analytics join at the query/page/day level. Flag cases where impressions are stable but clicks/sessions fall >15% over 90 days. Prioritize by business value. Set up daily SERP snapshots for those priority queries. Detect when AI Overviews or feature snippets appear and log cited sources. Prioritize queries where competitors are cited. Implement an LLM probe suite (20–200 queries depending on scope) and run weekly. Store responses and parse for brand mentions and competitor citations. Optimize pages for AI extraction:- Add explicit “Quick answer” summaries at the top. Include structured data (FAQ, HowTo, Dataset) and clearly tagged citations for your primary sources. Ensure content contains distinctive entity signals (authoritative names, research titles, unique statistics) that LLMs can map to your brand.
Thought experiments to sharpen strategy
Thought experiment 1 — The Invisible Citation: Imagine two pages, equal in rank and quality. One includes a short, clearly labeled 80‑word summary with a link to an internal dataset; the other buries the summary mid‑page. Ask: which is more likely to be selected as the single citation in an AI Overview? Run probes to test. The expected outcome: the concise, front‑loaded summary wins. This is low‑cost and high‑impact.
Thought experiment 2 — The Synthetic Query Bank: Assume a competitor invests in a published benchmark report that gets cited by LLMs. If you invest in a comparable proprietary dataset and publish it with rich markup and an ISBN/DOI, what happens to citation share? Simulate by releasing a small dataset and tracking LLM probes and SERP citations over 8 weeks.
Thought experiment 3 — Attribution vs. Proof: Suppose your content drives 25% of MQLs but leadership wants “direct clicks to revenue.” Can you design a hybrid approach where content includes exclusive gated assets tied to a short conversion funnel in the test region? Use that to produce direct, attributable conversions for proof while continuing to measure uplift in open traffic.
Final takeaways
Data shows the likely driver: extractive SERP features and LLMs are diverting clicks even when positions remain stable. The remedy is not to chase rankings alone but to adapt content for AI extraction, create signals LLMs can cite, and build experimental attribution to prove ROI. With modest technical investment (SERP snapshots, LLM probes, server‑side tagging) and targeted content changes (summaries, schema, datasets), mid‑market B2B teams can stem organic decline, regain visibility in AI Overviews, and produce causal evidence of value to skeptical stakeholders.
If you want, I can: (a) provide a reproducible SQL + Python pipeline for the GSC + GA4 join and CTR decay flags; (b) share a headless‑browser SERP snapshot script and LLM probe prompts; or (c) outline an 8‑week experimental plan you can run with your analytics team. Which would be most useful?