How we research and write this library
We start with a question people ask AI assistants, record the web searches the assistant runs to answer it, and plan one page per question. Each page is written from sources we have read, checked a second time, and published with its author, sources and update date.
- Observed means an assistant ran the searches and we recorded them. Modelled means our own estimate, and it is labelled.
- Every statistic links to the page we read it on, with the date we opened it.
- Our observed coverage is partial, and this page shows exactly how much.
- We do not guarantee rankings or citations.
The five steps from question to published page
Every page in the library goes through the same five steps, in this order.
Seed prompt
We start with a question a real person would type into an AI assistant, such as "what is AI SEO". Each seed is written in plain words and tied to one page we plan to publish.
Fan-out searches
We give the seed to AI assistants that can search the web, and record the searches they run to answer it. We tell Claude agents to run 6 to 10 searches, so their count is our setting. Gemini chooses its own count. These deeper searches are called query fan-out, and they show what the page has to cover.
Page plan
Each seed becomes a page plan: the question, the page type and the phrases the page should answer. The type is a definition, a how-to guide, a comparison, a business-impact page, an industry guide or a regional guide.
Sourced draft
We read the sources behind each claim and draft the page with the answer first. Every statistic comes from the page it cites, with the date we opened it.
Checks and publish
A second pass re-checks each statistic and link, and drops anything we cannot confirm. Then the page goes live with its author, its sources and an update date.
Live numbers
These numbers are counted from the library's data files each time the site is built. Nobody types them in by hand, so they move as the library grows.
- Planned pages
- 301
- Published pages
- 301
- Seed prompts
- 361
- Seeds with observed fan-out
- 42
- Observed searches in total
- 212
- Average searches per observed seed
- 5.0
- Engines used
- 2
Engines used so far, and how many seeds each one has covered:
- Claude with web search: 39 seeds
- Gemini with Google Search: 19 seeds
A seed prompt is one question we research. There is one for each planned page, plus 60 extra buyer prompts for industry pages. We have observed the searches for 42 of 361 so far.
Observed or modelled?
You will see two kinds of search phrases in this library. We keep them apart on purpose.
Observed
An AI assistant with web search ran these searches, and we recorded them word for word. They appear in the box titled “What an AI assistant searches when it answers this”, with the date and the engine.
We told Claude agents to research each question with 6 to 10 searches, so the number of Claude searches is our setting, not Claude's own choice. Gemini runs were left to choose how many searches to make.
Modelled
Our own expansion of the sub-questions a buyer is likely to ask. We build them from the observed pattern and from the topic. Each page labels them as modelled, and we never present them as a record of what an assistant did.
Observed coverage is partial
Search APIs have free-tier quotas, and an assistant can only run so many searches in one turn. We cannot observe every seed in one go, so we work through the seeds in passes. Each pass adds to the numbers above.
To extend the observed set, run these two commands from the project folder:
node scripts/aiseo/fanout.mjs --pass 1 node scripts/aiseo/merge-research.mjs
The first command records a new pass of searches and needs a Gemini API key in .env.local. The second merges every saved run into the file the site reads.
Examples of observed fan-out
These are the 8 observed seeds with the most searches. Open a row to read every search we recorded for it. They are a sample, not a rule.
| Seed prompt | Searches observed |
|---|---|
| “what are vector embeddings and semantic search for SEO”Page: What Are Vector Embeddings and Semantic Search?Engines: Gemini with Google Search, Claude with web search | 12 searches
|
| “how does Perplexity choose which websites to cite”Page: What Is Perplexity and How Does It Pick Sources?Engines: Gemini with Google Search, Claude with web search | 9 searches
|
| “what is AI SEO”Page: What Is AI SEO?Engines: Gemini with Google Search, Claude with web search | 8 searches
|
| “what is answer engine optimization AEO”Page: What Is Answer Engine Optimization (AEO)?Engines: Gemini with Google Search, Claude with web search | 8 searches
|
| “what is generative engine optimization GEO”Page: What Is Generative Engine Optimization (GEO)?Engines: Gemini with Google Search, Claude with web search | 8 searches
|
| “what is a knowledge graph and how does Google use it”Page: What Is a Knowledge Graph?Engines: Gemini with Google Search, Claude with web search | 8 searches
|
| “what is LLM optimization LLMO”Page: What Is LLM Optimization (LLMO)?Engines: Gemini with Google Search, Claude with web search | 8 searches
|
| “what is zero-click search and how does it affect traffic”Page: What Is Zero-Click Search?Engines: Gemini with Google Search, Claude with web search | 8 searches
|
Our sourcing rules
Facts are where trust is won or lost, so the rules for them are strict.
- Read the page, not a summary. Every statistic is read on the page it cites. We do not repeat a number we only saw quoted somewhere else.
- Dated. Each source shows the day we opened it, and each guide shows when it was last updated.
- Checked twice. A second pass re-reads each statistic against its source. If we cannot confirm it, we drop it.
- Listed on the page. Sources appear at the bottom of every guide, with the claim each one supports.
- Primary sources first. We prefer a platform's own documentation or the original study over articles that report on them.
- No number without a source. If we cannot find a reliable figure, we say so instead of estimating. This is why the library has no search-volume figures.
- Other firms' results stay theirs. Results published by other firms keep their attribution and are never shown as Chalk Labs work.
Editorial standards: E-E-A-T in practice
Google's quality guidelines describe E-E-A-T: experience, expertise, authoritativeness and trust. Here is what each one means for this library. For a plain explanation, read What Is E-E-A-T and Why Does It Matter for AI Search?.
Experience
Chalk Labs runs this method on its own site and on client sites, and publishes those results with their limits. See the case studies and the results on Rahil's page.
Expertise
Every guide has a named author, Rahil Jain, and a written method, which is this page. His verified background is listed on his page, and nothing beyond it is claimed.
Authoritativeness
We cite primary sources first, such as a platform's own documentation or the original study. Every claim links to the page where we read it.
Trust
Chalk Labs sells AI SEO services, so we have a commercial interest, and each guide says so. Guides are drafted with AI research assistance, and each guide says that too. Every guide shows when it was published and last updated, and gives an email address for corrections.
Corrections and updates
Found a mistake? Email chalk.core@gmail.com with the page address and what is wrong. We aim to fix a confirmed factual error within 7 days. When a page changes in a meaningful way, its update date changes too.
Dates show in four places:
- Under the title of every guide, as “Updated”.
- In the note at the bottom of every guide, which gives the published date and the last updated date.
- Next to every source, as the day we opened it.
- In each guide's structured data, as the published and modified dates.
Limits of this research
Good research says what it cannot do. These are the limits of ours.
- Assistants differ and they change. A search pattern we saw on October 20, 2026 may not repeat next month, or on another assistant.
- Observed samples are small. We have observed 42 of 361 seed prompts, with about 5.0 searches each. A small sample can mislead.
- Some engines are not recorded yet. Our observed searches come from Claude with web search and Gemini with Google Search. We have no recorded fan-out from ChatGPT, Perplexity, Google AI Mode, Google AI Overviews and Microsoft Copilot, and they may behave differently.
- We do not publish search volumes. We could not source them from a primary provider, so we leave them out instead of estimating.
- Fan-out shows what an assistant searched, not how many people ask. A phrase an assistant uses may be rare in real life.
- A statistic is true as of the day we opened its source. Numbers go stale, which is why every source shows its date.
- Following the method does not guarantee a citation or a ranking. Search engines and AI models are run by third parties, and we do not guarantee outcomes.