An AI answer is assembled in six layers: what the model learned in training, the product and context around it, the plan for which searches to run, the search results it retrieves, the pages it reads and selects, and the final answer with its citations. A brand can influence and observe some layers far more than others. Search rankings feed one layer; they do not decide the answer.
This post is part of Zia Wasi's GEO (generative engine optimization) research at Depra AI. Zia Wasi, co-founder of Depra AI, collected it from public provider documentation, checked on 13 and 14 October 2026, alongside the Hinglish vs English AI shopping study. Read it alongside the Depra AI methodology, which explains how Depra AI captures and scores each engine. The layers are an analytical model, not a design any engine has published.
What are the six layers of an AI answer?
The six layers are model knowledge, product and context, query planning, retrieval, reading and selection, and answer and action. An engine can combine, skip or repeat them. An answer from memory skips retrieval; a hard question may loop through it several times.
| Layer | What happens | What a brand can observe |
|---|---|---|
| 1. Model knowledge | Patterns the model learned in training. No live lookup is needed. | Brand mentions in tests run with search off. Which training data taught the model is usually unknown. |
| 2. Product and context | The app, its instructions, the conversation, memory, location and any connected data shape the task. | The product, mode, account state, country and language the question was asked in. |
| 3. Query planning | The system decides whether to search and can turn one question into several searches. | The searches it ran, where the product shows them. |
| 4. Retrieval | A search index, a shopping or social feed, or a connected tool returns candidate pages. | Returned URLs and their positions, where an API exposes them. |
| 5. Reading and selection | The system fetches and filters the candidates before it writes. | Pages consulted, where exposed. Not every candidate is read. |
| 6. Answer and action | The model mentions, cites, compares or recommends a brand. The buyer may click or buy. | The answer text, its citations, the tone around each brand, and referral visits. |
Do Google rankings decide what AI answers cite?
Rankings matter because they decide which pages enter the candidate pool. They are not enough, because filtering, reading and writing happen after retrieval, and each step can drop a page that ranked well.
Google states the entry rule plainly. To appear as a supporting link in AI Overviews or AI Mode, a page "must be indexed and eligible to be shown in Google Search with a snippet", and there are no extra requirements (Google Search Central, AI features, checked 14 Oct 2026). The same page says both features may use "query fan-out": several related searches across subtopics. So the page that gets cited may rank for a sub-question, not for the question the buyer typed. Our query fan-out guide shows the real searches ChatGPT ran for shopping questions asked from India.
OpenAI draws the second line. Its web search tool returns a list of every source the model consulted, and the page notes that the number of sources is often greater than the number of inline citations (OpenAI web search guide, checked 14 Oct 2026). A page can be read and still not be cited.
Perplexity runs its own index. It announced a search index of "billions of webpages" when it opened its Search API on 25 September 2025 (Perplexity, checked 14 Oct 2026). A Google ranking does not show where a page sits in that index.
What is the difference between grounding and training?
Grounding means giving the model outside evidence at the moment it writes an answer. Training is what the model learned before release. A new page can help a searching assistant long before it could reach any future training set, and a model can name a familiar brand without opening any current page.
Gemini shows the grounding loop in its own documentation: the model reads the prompt, decides whether a Google Search would improve the answer, writes one or more queries, runs them and then answers (Gemini API, grounding with Google Search, checked 14 Oct 2026). OpenAI also separates the two at the crawler level. OAI-SearchBot is the crawler OpenAI uses to surface sites in ChatGPT's search features, GPTBot gathers content that may be used to train its models, and each setting is independent of the others (OpenAI crawlers, checked 14 Oct 2026). Blocking GPTBot does not remove a site from ChatGPT search. The AI crawlers list covers every bot.
Connected tools, such as an MCP connection, help only the people who connected them. They put nothing into a public search index.
What is the difference between a mention, a citation and a recommendation?
A mention is the brand name in the answer text. A citation is a link to a web page that supports the answer. A recommendation is an answer that presents the brand as a good fit for what the buyer asked. Each measures a different thing, and one can happen without the others.
- Mention without citation: the model names the brand from training or from a page it read but did not link.
- Citation without mention: the answer links a marketplace or review page that lists the brand, but the text never names it.
- Mention without recommendation: the answer names the brand in a warning, a comparison it loses, or a list of options to avoid. A warning is not a recommendation.
- A citation is not a visit: a cited page only earns traffic if the buyer clicks. Referral visits are a separate number.
What should a brand measure in AI answers?
Measure each layer with its own number, and show the sample size, date range and engine beside every percentage. Every rate counts successful answers only, so an outage never looks like lost visibility.
- Mention rate: answers that name the brand, divided by all eligible answers. Keep questions that already contain the brand name in a separate group.
- Recommendation rate: answers that present the brand as suitable for the buyer's request, divided by eligible answers.
- Citation rate: answers that cite a page on the brand's own site, divided by eligible answers. Track outside pages that discuss the brand separately.
- Search usage rate: runs where the engine searched, divided by runs where search was allowed. Unknown is not the same as no search.
- Retrieval coverage: search checks where an owned page appeared within the result depth inspected, divided by valid checks. State the depth.
- Measured recommendation share: the brand's recommendations divided by all brand recommendations in the sample. This is share of a sample, not market share.
Two rules that keep the numbers honest
- Use rank only when the answer actually ranks options. The order in which prose names brands is not always a preference order.
- Never add or average numbers with different denominators, such as a platform's own citation counts and sampled answers, into one figure.
Which claims should an AI visibility tool avoid?
A tool should avoid any claim that skips a layer. The evidence supports none of the claims below.
- IndexNow indexes a page in every AI system: it notifies participating search engines, each engine makes its own decision, and submission does not guarantee indexing (IndexNow FAQ, checked 14 Oct 2026). Do indexing requests get you into AI answers? covers what each request does.
- A missing llms.txt file makes a site uncrawlable: it does not. Does llms.txt work? covers what the file does.
- Allowing training bots produces recommendations: training and search are separate routes, as OpenAI's own crawler page shows.
- A first-page Google position guarantees a citation: selection happens after retrieval.
- A citation is a customer visit, or a sampled share is market share: both are different numbers.
- "Engine X always prefers site Y": no universal preference is verified. Source priorities should come from real buyer questions and the citations they return.
What does this look like in real AI shopping answers in India?
The retrieval layer moves between waves, so the sites an engine cites are a measurement, not a rule. In the August 2026 wave of the Hinglish vs English AI shopping study (480 answers from ChatGPT, Gemini and Perplexity), reddit.com was cited in 106 answers, youtube.com in 100, nykaa.com in 93, and amazon.in and flipkart.com in 79 each.
In the October 2026 re-run (479 answers, same design), amazon.in was cited in 132 answers, flipkart.com in 82, hindustantimes.com in 76, nykaa.com in 65, thedermaco.com in 64 and myntra.com in 44. The Perplexity brand gap between English and Hinglish questions was 28.2 points, against 2.0 for Gemini and 0.8 for ChatGPT. Both waves are on the study page, and the August answers, prompts and analysis code are public on GitHub.
How does Depra AI label each run by engine?
Depra AI stores every answer with its engine, country, language and date, and reads it at the answer layer. Each engine is reported on its own, because each assembles answers differently. Even within Google, AI Mode, AI Overviews and the Gemini app can give three different answers.
- Mentions: each needs a verbatim quote and at least 50% confidence to count, with its position and tone recorded, so a warning reads as a warning.
- Branded questions: prompts that contain the brand name are left out of scores by default.
- Sources: every cited website is labelled by site type, marketplace included, which feeds citation gap analysis and marketplace intelligence.
- What it cannot see: pages an engine read but did not cite, and ChatGPT answers with search off.
| Engine | Capture route | Extra layer data stored | How often |
|---|---|---|---|
| ChatGPT | chatgpt.com web app, web search on | The web searches ChatGPT ran (query planning), and the products it showed (the Shopping report, on Growth, Scale and Enterprise) | Daily on every plan |
| Gemini | gemini.google.com web app | Cited pages | Daily on every plan |
| Google AI Overviews | A real Google results page in that country | Whether an Overview appeared at all, stored as its own result | Daily on every plan |
| Perplexity | Sonar, its answer model with live web search | Cited pages | Weekly on Starter, Growth and Scale; daily on Enterprise |