Lesson 0005 · Did it work?
Measuring It
Two scoreboards, wildly asymmetric. Classic search hands you a first-party API. AI citations hand you nothing — so you build the proxy everyone else builds.
Recap from Lesson 0004: you made a page eligible (0002), labelled its facts (0003), and shaped its prose for retrieval (0004). All of that is a bet. This lesson closes the loop — how do you know it landed?
The two pipelines don’t measure the same way, and the gap is the whole story. For classic search, Google gives you ground truth through an API. For AI-answer citations, no engine exposes one — so every tracker on the market, and the one you’ll build, is a proxy.[1]
citation_share.py on a set of answer-engine responses and get your coverage (% of prompts that cite you) and share of voice (your slice of all citations) vs competitors — the same proxy metric Profound and Ahrefs Brand Radar sell. Two scoreboards
Google Search Console API. Real impressions, clicks, average position, and the exact queries — from Google's own logs.[2]
The Web search type now folds in traffic from AI features, so AI-Overview clicks already show up here.
✓ ground truth, queryable, free
No OpenAI / Anthropic / Perplexity / Google endpoint answers "how often is my site cited?" Google's June 2026 gen-AI reports are a Console UI view, not in the API yet.[3]
So you sample: run a prompt library, parse the citations, tally.
✗ proxy only — synthetic, not real traffic
Classic side: just call the API
This one’s a solved problem — authenticate, then POST a date range and the dimensions you want. No tool to build; the platform is the tool.
# Search Console: top queries by clicks, last 28 days
POST https://www.googleapis.com/webmasters/v3/sites/{siteUrl}/searchAnalytics/query
{
"startDate": "2026-05-20", "endDate": "2026-06-17",
"dimensions": ["query"],
"rowLimit": 100
}
# -> rows of { keys:[query], clicks, impressions, ctr, position }Poll it on a schedule, store the rows, and you have rank/impression/click trends over time — the backbone of any SEO monitor.
AEO side: the proxy you build
Because there’s no ground truth, every commercial AEO tracker does the same three steps — and so does tools/citation_share.py:
| Step | What it is | Who owns it |
|---|---|---|
| 1 · Prompt library | A fixed set of buyer questions in your space | you (the API key + prompts) |
| 2 · Run the engine | Send each prompt, capture response + cited links | you (one API call) |
| 3 · Parse & score | Extract cited domains → coverage + share of voice | citation_share.py |
A second helper, tools/build_prompts.py, owns step 1: it assembles that buyer-question prompt library — from an LLM fan-out or your Search Console rows — in the exact shape citation_share.py consumes. The tool is the offline, testable core. It scores cited links (the AEO equivalent of a rank), not bare brand mentions:
# coverage = prompts that cite me; SOV = my citations / all citations
cited = cited_domains(entry) # links in response + citations[]
hit = bool(cited & my_domains)
sov = mine / total_citations- Self-check (offline):
python3 tools/citation_share.py --demo - Hand-make a tiny
results.json— 3–4 buyer questions, paste in real answers + links from ChatGPT/Perplexity — then score it:python3 tools/citation_share.py results.json --domain yourdomain.com - Re-run it next week. The single number that matters is your own coverage trend — not the absolute value.
$ python3 tools/citation_share.py results.json --domain myseo.io AI-citation share — 4 prompts, you = myseo.io ──────────────────────────────────────────────────── COVERAGE 50% cited in 2/4 prompts SHARE OF VOICE 33% of all citations are yours ──────────────────────────────────────────────────── Top cited domains: 3ahrefs.com 2myseo.io ← you 1moz.com ──────────────────────────────────────────────────── Per prompt: [HIT ] best seo audit tool for developers? [miss] how to track rankings via api? [HIT ] what is generative engine optimization? [miss] cheapest backlink checker?
Ceiling to know: citation_share.py counts cited links only. “Brand mentioned without a link” is a different (softer) metric; domain matching is naive (no public-suffix list). Both are noted in the tool’s docstring as upgrade paths.
Retrieval practice · no peeking
Read the scoreboard
Answer from memory — that effort is what makes it stick. One try each; pick before you read the others.
The exact request body and the rows you get back — your classic-search ground-truth feed. Start at the <a href="https://developers.google.com/webmaster-tools/about">API overview</a>. For the AEO side, see the proxy-method walkthrough in <a href="https://ahrefs.com/blog/chatgpt-visibility-tracking/">Ahrefs' ChatGPT-visibility guide</a> and the gaps note in <a href="/resources/"><code>RESOURCES.md</code></a>.