More people now ask an AI assistant before, or instead of, searching. For brands and researchers that raises a new question: what do ChatGPT, Perplexity and Google AI Mode actually say about a topic, and which sources do they cite? Answering it means collecting answers the way users see them and analyzing them with more care than a ranking.
Why AI answers are a different kind of data
A search result page is a list with positions. An AI answer is generated text that may mention brands, recommend options and cite sources. Three properties shape how you collect it:
- Variability. The same question can produce different wording, different recommendations and different citations from one run to the next.
- Context sensitivity. Location, language, conversation history and the product surface all influence answers.
- Partial sourcing. Some answers cite many sources, some cite a few, and some cite none. A missing citation is not proof that a source was unused.
Because of variability, a single answer is an anecdote. Reliable conclusions come from repeated observations of a fixed set of questions.
How the three engines differ
| ChatGPT | Perplexity | Google AI Mode | |
|---|---|---|---|
| Answer style | Conversational, often structured | Concise, research-oriented | Search-integrated overview |
| Citations | Present when search is used; varies | Usually several sources | Linked source cards |
| Typical use in research | Brand mentions and recommendations | Source visibility | How Google's AI frames a topic |
| ProtoScrape credits per success | 3 | 3 | 2 |
These are tendencies, not guarantees; each product changes frequently. Google AI Mode is also distinct from AI Overviews on the classic results page. See AI Mode vs AI Overviews vs SERP.
Design a study before collecting
- Fix the questions. Write the prompts once and reuse them exactly, grouped by intent such as “best tools”, “comparisons” and “how to”.
- Repeat on a schedule. Daily or weekly runs show stability and trends instead of one-off answers.
- Separate the measures. A mention in the answer text, a citation of your domain and a recommendation are three different metrics.
- Report sample sizes. “Mentioned in 14 of 20 answers” is honest; “ranks first in ChatGPT” is not a meaningful claim.
- Keep failures out of the denominator. A failed collection is missing data, not an answer without your brand.
Collect all three with one API
ProtoScrape uses the same job API for every engine, so a study can submit identical questions to each and compare normalized results: answer_text plus citations with titles and URLs.
for engine in chatgpt perplexity google_ai_mode; do
curl https://api.protoscrape.com/v1/jobs \
-H "Authorization: Bearer $PROTOSCRAPE_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: crm-question-$engine-2026-11-05" \
-d "{\"engine\":\"$engine\",\"query\":\"Which CRM is best for a five-person team?\"}"
done
For larger studies, a batch submits up to 100 prompts at once, and scheduled collections repeat them daily or weekly. Results export to CSV or JSON, and the workspace's Result preview renders each answer with its sources so you can read it as a user would.
Read answers critically
An AI answer can be confidently wrong, outdated or inconsistent with its own sources. Treat citations as what the system displayed, not as proof that each source supports each claim. When a result matters, open the sources and check.
The AI visibility tracking guide turns these observations into a repeatable report, and the citation analysis guide shows how to compare sources across engines. ProtoScrape is independent and is not affiliated with OpenAI, Perplexity or Google.