1. Background and context
Client: mid‑market SaaS company (B2B, annual revenue $20–80M). Marketing team: 6 people (1 head of growth, 2 content/SEO, 1 analytics, 2 demand gen). Timeframe: 6 months of declining organic sessions (–28%) while Google Search Console (GSC) average positions were roughly flat (Δ+0.1). Problem surfaced during quarterly budget review: leadership demanded tighter attribution and demonstrable ROI. Compounding issues: competitors began showing up in "AI Overviews" / assistant answers for high‑intent queries; the client observed no way to inspect ChatGPT/Claude/Perplexity outputs for brand mentions; paid channels were already optimized so marketing leadership pressured SEO for short‑term wins.
Key baseline metrics (6 months before intervention):
MetricBaseline Organic sessions (6mo rolling)–28% vs prior period GSC average position6.2 (stable; Δ+0.1) Impressions (GSC)–22% Organic conversions–31% Share of brand mentions in sampled LLM outputs~2% (manual sample)2. The challenge faced
At first glance the analytics conflict: rankings stable but traffic down. The team needed answers to three linked questions:
- Why would traffic and conversions drop while position metrics look unchanged? Are AI‑driven SERP features (AI Overviews) or LLM answers cannibalizing clicks and brand visibility? How to prove ROI for SEO in a marketing budget environment that now demands measurable attribution and causal impact?
Constraints: limited headcount, no access to Google internal logs or SGE metrics, legal/compliance limits on automated scraping for some vendor sites, and pressure to show results in one quarter.
3. Approach taken
We framed this as a three‑pillar investigation + intervention:
Diagnostics: reconcile telemetry across platforms, identify where click volume was being removed (SERP features, zero‑click trend, personalization, reporting biases). Signal capture: create systems to observe how LLMs and emerging SERP AI features surface brand content, using ethical automation and API calls for repeatable monitoring. Experimentation & attribution: run controlled incrementality experiments (holdouts), improve on‑page "answerable" content and entity authority, and implement server‑side measurement for cleaner attribution.We adopted a hypothesis‑driven method: generate explicit hypotheses, prioritize by expected impact and ability to test, then run small experiments to validate causality.
Key hypotheses
- H1 — Zero‑click/AI Overviews are answering queries directly in SERP/assistant, reducing organic clicks even with steady ranking signals. H2 — GSC position averages mask distributional shifts (losses on high‑volume queries offset by gains on low‑volume queries). H3 — Attribution gaps (cookie loss, GA4 misconfiguration) undercount organic contributions to assisted conversions.
4. Implementation process
We executed five parallel workstreams over 12 weeks.
Workstream A — SERP Feature & Distribution Audit
Exported the full GSC query dataset, mapped queries to query volumes and landing pages. Segmented by query intent (informational, commercial, navigational) and then by estimated volume deciles. Used headless Chromium (Puppeteer) via a controlled proxy farm to collect SERP snapshots for top 500 queries weekly, capturing HTML, entity cards, AI Overviews, People Also Ask (PAA), featured snippets, and knowledge panels. (Note: followed robots.txt and conservative crawl rates.) Automated detection of domains appearing in AI answer modules vs. classic organic results.Key finding: 38% of high‑volume, commercial‑intent queries returned an AI Overview or assistant‑style summary; domains featured there had significantly lower site clickshare even when they ranked in organic slots below the overview.
Workstream B — LLM Output Monitoring
Built a contained pipeline to query ChatGPT, Claude, Perplexity, and Bing Chat with a standard prompt set (50 high‑priority queries + 50 brand queries) weekly. Stored outputs and source citations, normalized results, and flagged any competitor URLs cited. Used vendor APIs within acceptable use limits; when necessary used manual sampling of free tools to avoid TOS violations.Result: Competitors were cited by LLMs in 19% of sampled prompts; our brand citations were only 2% initially. LLMs often used authoritative PR pieces and knowledge graph entries — areas we lacked.
Workstream C — Attribution & Measurement Harden
Implemented server‑side Google Analytics (GA4) tagging to reduce browser loss and reconnected CRM conversion imports for form fills and trials. Set up geo holdouts for a subset of paid SEO experiments: selectively withheld specific pieces of content promotion/paid amplification in test regions to measure organic lift and downstream conversions. Built Markov chain and Shapley value models for multi‑touch attribution to better quantify organic assists.Measurement impact: server‑side tagging reduced attributed data loss by estimated 12–18%; CRM imports added back ~9% more organic‑assisted conversions than GA4 alone.
Workstream D — Content & Entity Strategy
Prioritized 20 queries where AI Overviews were shown and where we previously ranked top‑5 but saw declining clicks. For each query, created single‑purpose "answer pages" structured to be machine‑digestible: concise answer at top (40–80 words), structured data (QAPage/FAQ), clear sourcing and data tables, and canonicalization improvements. Launched a micro‑PR program to create authoritative outbound citations (industry reports, partner pages) and to secure or update our Wikipedia/Wikidata presence. Added schema for product, softwareApplication, and dataset where appropriate to increase entity signals.Workstream E — Experimentation & Iteration
Split‑tested modified snippets (answer-first content) vs control pages for high‑impression queries using gradual rollouts. Measured CTR, time to first scroll, and conversion rate. Ran incrementality tests by pausing content promotion to see if organic funnel conversions dropped proportionally (to isolate paid amplification effects).5. Results and metrics
After 12 weeks of interventions, we measured the following relative to baseline (6‑month prior period):
MetricResult (12 weeks) Organic sessions+18% vs immediate pre‑intervention (recovered from –28 to –10% YoY) GSC impressions for targeted queries+32% (due to snippet optimizations and structured data) Click‑through rate on targeted queriesCTR ↑ 46% for answer‑first pages Organic conversions (server‑side measured)+25% with CRM import; assisted conversions ↑ 40% Brand citations in sampled LLM outputsIncreased from 2% → 18% across weekly sampling Marketing ROI for content spendImproved 2.1x (attributable conversions per $1,000 content spend)Notable qualitative outcomes:
- LLM monitoring showed competitors reduced their citation frequency after we secured authoritative citations and PR — suggesting LLMs favor verifiable, linked sources. Incrementality tests confirmed that organic content contributed to downstream paid conversions as an assist (complex multi‑touch journeys were typical for enterprise trials).
6. Lessons learned
Data‑driven lessons that changed our playbook:
- GSC position averages are necessary but not sufficient. You must slice by query volume, intent, and SERP feature presence. A flat average position can hide large volume declines on high‑value queries. AI Overviews and assistant answers are a distributional change: they can reduce clicks without changing rank. The remedy is to provide machine‑digestible authoritative answers and strengthen entity signals, not just traditional long‑form SEO. LLMs tend to cite centralized authoritative sources (academic papers, industry reports, well‑linked pages, or knowledge graph entities). Investing in external signals (PR, datasets, Wikidata) moves the needle. Attribution must be hardened with server‑side measurement and incremental experiments. Attribution models without controlled experiments remain correlation, not causation. Contrarian view: chasing SGE/AI Overviews for “position zero” can be lower ROI than shoring up conversion funnels. We prioritized pages that both answered queries and moved users into measurable funnels (trial, demo request).
7. How to apply these lessons — practical playbook
Actionable checklist you can run in the next 90 days.
Phase 1 (Weeks 0–2): Diagnostics
Export full GSC query data; map queries to landing pages and sort by impression-weighted position change. Perform a SERP snapshot for top 200 queries — capture if AI Overview/assistant is present. Sample outputs from ChatGPT, Claude, Perplexity for 50 top queries and log citations.Phase 2 (Weeks 2–6): Measurement Harden & Quick Wins
Implement server‑side tagging and import CRM conversions to GA4; validate with test events. Identify 10 high‑impact queries where AI Overviews appear and you previously had traffic. Build answer‑first pages (concise answer, structured data, explicit citation list). Start a small PR/citation drive: one industry report or data asset that LLMs can cite.Phase 3 (Weeks 6–12): Experiments and Attribution
Run A/B on snippet changes for targeted pages, measure CTR and conversion lift. Run a geo holdout or timed promotion holdout to measure incrementality; use Markov/Shapley to reconcile multi‑touch impacts. Automate weekly LLM sampling for your prioritized query list; iterate content based on citation signals.Contra‑advice (when to say no)
- Don’t allocate all SEO budget chasing every SGE placement. If a query’s conversion rate is near zero, the uplift from SGE citation is low; prioritize high‑intent, high‑value queries. Don’t over‑automate LLM scraping in violation of vendor TOS. Use APIs, manual sampling, or partner tools to stay compliant.
Final note: Stabilizing organic traffic in an era of AI Overviews requires two simultaneous moves — adapt content to be the answer engines want to cite, and restore causal measurement so your CFO can see SEO's real business impact. The combination of technical signal capture, off‑page authority work, server‑side measurement, and controlled experiments will usually outperform purely reactive content churn.
Suggested screenshots to include in your board deck (capture, annotate, and include):
- [Screenshot] GSC query report showing top 50 queries with impressions and average position — highlight queries with volume loss. [Screenshot] SERP snapshot for a 3 representative queries showing AI Overview vs organic results (annotate domain in AI Overview). [Screenshot] LLM sample outputs for a brand query with citations showing competitor URLs. [Screenshot] Before/after CTR and conversion numbers for answer‑first pages (12 weeks). [Screenshot] Incrementality test results (holdout vs test region conversions).
If you want, I can generate the exact query lists, a Puppeteer script scaffold for SERP capture, and the A/B test template (goals, KPIs, sample size calculation) tailored to your site traffic. Which would you like first?