Organic sessions are down. Google Search Console (GSC) shows rankings are stable. Competitors who historically weren't visible are now showing up in AI Overviews (ChatGPT/Claude/Perplexity) and you can't see what those models say about your brand. Marketing budget is under scrutiny and stakeholders demand clear attribution and ROI. How do you diagnose, measure, and respond — with data https://jsbin.com/wejoxucixa — not guesswork?
1. What you'll learn (objectives)
- How to determine whether traffic loss is real, an attribution artifact, or caused by AI-driven consumption (AI Overviews / chat assistants). How to probe and capture what large language models (LLMs) and AI Q&A services say about your brand and competitors. How to run tests and instrument your site to prove attribution and ROI to stakeholders. Practical, repeatable steps to improve the probability that models cite or excerpt your content (increasing "AI visibility"). How to design lift and incrementality experiments to show budget impact with statistical rigor.
2. Prerequisites and preparation
Before you start, gather this baseline data and access:
- Access to Google Search Console and permission to export queries/impressions/clicks for the last 90+ days. Analytics access (GA4 preferred) plus BigQuery export or another raw-event store. Server logs or a way to access raw request logs for at least 30 days. CRM access (HubSpot, Salesforce) to connect leads back to landing pages and timestamps. Ability to create A/B pages or experiment variants and add UTM parameters/unique phone numbers/promo codes. Tools: a SERP API (SERP API, Google SERP), a model-query API (OpenAI/Anthropic/Perplexity), and a rank tracker.
Preparation checklist
Export GSC data by page, query, country, and device for the previous 90 days. Extract page-level session trends from GA4 and server logs for the same period. Pick a representative set of 10–30 pages that lost traffic but kept rank position according to GSC. Create a simple observability spreadsheet or BigQuery table to compare impressions → clicks → sessions → conversions.3. Step-by-step instructions
Step A — Confirm the problem: is this real traffic loss or attribution noise?
Compare channels: is organic sessions down while clicks from GSC are flat? Export impressions vs clicks vs CTR vs sessions by page. Do numbers diverge per page? Check server logs: are organic requests (HTTP referrer contains google) down? Server logs reveal actual page requests independent of GA/GSC artifacts. Capture full referrer chains where possible. Look for analytics breakages: did you recently deploy GA4 changes, consent banners, or tag manager changes? Run a quick real-time test with a private page visit and confirm event arrival in GA4 and logs. Check search features: are impressions moving from clicking-through results to zero-click features like featured snippets, People Also Ask, or AI Overviews? Use SERP API snapshots for your target queries and compare feature presence over time.Step B — Measure "AI leakage": how much demand is being satisfied by LLMs and answer engines?
Pick representative queries for your pages. For each query, capture the top-10 organic results plus AI Overviews outputs at multiple times of day (use an API to avoid rate limits). Programmatically query ChatGPT, Claude, and Perplexity with the same search prompts. Save outputs and any cited sources. Ask: does the model cite your domain? If yes, how often and with what excerpt? Record how often competitors are cited versus you. Are competitors being referenced despite lower SEO metrics? Take screenshots of AI outputs and SERP features for documentation. Store timestamps for later correlation with traffic shifts.Step C — Add measurement hooks to capture AI-driven conversions
LLMs rarely provide referrers. So measure downstream user behavior:
- Add or rotate unique, short-lived UTM-tagged landing pages that you can test in content likely to be surfaced by models (concise "answer" pages, FAQs, TL;DRs). Deploy unique phone numbers or promo codes on those pages, and reserve them for the experiment so incoming leads can be attributed even if no referrer is present. Log the first page a user lands on (server-side) and pass that into the CRM so you can tie leads back to initial exposure even if later touchpoints are direct.
Step D — Run incrementality tests to prove ROI
Choose a pool of comparable pages and split them into control and treatment groups (randomized by topic/traffic). Treatment group: add AI-friendly summary content, structured data, citation-ready snippets, and machine-readable TL;DRs. Run the test for 4–8 weeks. Measure lifts in untracked leads (unique phone numbers/promo codes) and tracked conversions (UTM + CRM match). Use BigQuery to run cohort analysis. Compute statistical significance and incremental revenue. Present results as lift in qualified leads per 1,000 impressions and the associated CPL (cost per lead) using your current spend baseline.4. Common pitfalls to avoid
- Don't assume GSC clicks = sessions. GSC filters queries differently and aggregates differently from analytics. Don't treat LLM outputs as static — they change. Capture timestamps and multiple samples per query. Don't over-optimize to "chat-bait" phrasing in a way that reduces content utility for human readers. Aim for both human and model-friendly clarity. Avoid relying solely on surface SEO metrics (Domain Authority, 'SEO score') to judge visibility. LLMs may prioritize concise, well-cited authority for specific answers. Don't use a single attribution model. Use last-click, multi-touch model, and—most importantly—incrementality tests to argue causality.
5. Advanced tips and variations
Make your content AI-ready (but human-first)
- Provide a one-paragraph summary at the top of pages ("TL;DR") with a clear answer to common queries. Models often excerpt short concise answers. Structure content into question-and-answer blocks with clear H2/H3 headings and short (40–120 word) definitive answers. Use schema.org QAPage, FAQPage, and Answer markup where relevant. Include succinct, citation-ready statements with data and links to studies or primary sources. Models favor content that includes explicit citations and authoritative sources.
Improve your signals of trust and authority
- Push verifiable metadata: structured data for Organization, logo, sameAs (Wikidata, LinkedIn, Crunchbase), and Claim Knowledge Panel where possible. Get high-quality citations: reach out for editorial links, guest posts on domain-authoritative sites, and primary research publications. LLMs often surface these signals.
Programmatic monitoring of LLM outputs
- Automate daily sampling for target queries across model providers and store outputs in a database. Tag outputs for brand mention, sentiment, and citation source. Use simple NLP to compute "brand share" in answers over time (percentage of responses that mention your brand vs competitors).
6. Troubleshooting guide
Symptom: Rankings are stable in GSC but sessions dropped 30%
Check snippet features: Did an AI Overview or featured snippet start appearing for your queries? Use SERP API to confirm and find who it cites. Verify click-through changes: Has SERP layout changed (more PAA boxes, image blocks, or news carousels)? These reduce CTR even when rank is stable. Look for indexing cannibalization: Are multiple pages on your site competing for the same query with different meta descriptions? Consolidate canonical signals. Review cookies/consent changes: If a consent banner blocks analytics before pageview, sessions will be undercounted while server logs show actual visits.Symptom: Competitors appear in AI outputs but they have worse SEO scores
Examine the competitor's content for concise, answer-first copy or unique data. Models reward succinct, well-cited content more than overall SEO authority for some queries. Check if competitors published new research, case studies, or datasets that are easily citable. Unique primary data = high chance of being referenced. Look for non-SEO distribution channels (press, Reddit, StackOverflow, academic citations) that increase the chance of being included in model training corpuses or retrieval indices.Symptom: Can't see what ChatGPT/Perplexity say about brand
Run controlled prompts: "Summarize options for [topic], and list sources used." Ask the model for citations and then evaluate whether your domain appears. Repeat across model providers. Use the models' citation-enabled products (Perplexity shows sources; newer chat models are being built with source citations). Capture and store outputs. If models don't cite sources, use public web archives and date filters to see if your content existed before the model's training cutoff or is present in the retrieval index the model uses.Tools and resources
- Analytics & data: Google Search Console, GA4 + BigQuery, Server logs (NGINX/Cloudflare logs), CRM (HubSpot, Salesforce). SERP & rank tracking: SERP API, Ahrefs/SEMrush, RankRanger, Sistrix. LLM & AI sampling: OpenAI API, Anthropic/Claude API, Perplexity.ai, Bing Chat (API or Edge automation). Experimentation and analytics: Looker Studio, R or Python (pandas, statsmodels), Bayesian A/B tools (e.g., CausalImpact, scikit-learn), Split or Optimizely for page variants. Content & schema tools: Schema.org docs, Google Rich Results Test, Structured Data Linter, Surfer SEO, MarketMuse.
Final expert-level insights and closing questions
Here's the unconventional angle: don't fight models by chasing keyword spikes — design your content to be the canonical, citable answer for a narrow set of high-value queries, then measure the downstream business impact with controlled experiments. LLMs often select concise, well-sourced answers. That means you can win not by outranking competitors in a classical sense, but by becoming the most trusted, most-citeable source for the exact questions buyers ask.
Questions to consider right now:
- Which 10 queries drive the most qualified leads, not just impressions? Can you create a one-paragraph, citation-backed answer for each of those queries and instrument it with unique measurement hooks? How many of the competitors cited by AI models are citing unique data you can match or beat with your own primary research?
Proof-focused next step: run a 6-week, randomized experiment where half your high-value pages get AI-optimized TL;DRs, structured data, and unique conversion hooks; the other half remain as-is. Measure lift in qualified leads and cost per lead. That lift is the kind of evidence finance wants when budgets are negotiated.
Need a templated audit, query script for ChatGPT/Perplexity, or BigQuery query to connect page sessions to CRM leads? Ask and I’ll provide ready-to-run artifacts (including screenshot prompts and sample SQL) so you can start the experiment this week.