ChatGPT, Gemini, Perplexity, Claude: how each AI engine picks its sources
Ask four AI engines to recommend an accountant in Chatswood and you can get four different shortlists. Each engine retrieves sources its own way, trusts different signals, and cites with a different temperament. If you only optimise for one, you stay invisible on the other three.
Here is how the four engines your prospects actually use differ, based on what we see running repeated queries for Australian accounting firms. Engine behaviour shifts with every product update, so treat this as a mid-2026 snapshot; the measurement method at the end is what stays reliable.
ChatGPT: two modes, and you need to win both
ChatGPT answers in one of two modes. For questions it considers stable, it answers from training data: what it absorbed about the web the last time it was trained. For local and current questions, it runs a live web search and reads a handful of results before answering.
The two modes reward different work. Memory-mode answers favour firms mentioned consistently across the web over years: directories, articles, reviews, association listings. Search-mode answers favour firms whose pages rank for the query and state extractable facts. OpenAI crawls with separate bots for each purpose, which is why our own robots.txt welcomes both GPTBot (training) and OAI-SearchBot (live search).
What it rewards: breadth of corroboration across the web, plus pages that answer the question directly when it searches.
Google AI Overviews and Gemini: the search index wears a new face
AI Overviews sit on top of ordinary Google results, and Gemini grounds its answers in the same index. For a local service query, that means the machinery accountants already know still applies: local rankings, your Google Business Profile, review volume and recency, consistent name and address details.
The difference is what happens after retrieval. The AI layer reads the top results and composes an answer, so ranking third no longer guarantees a click. Your page also has to be the one the summary quotes. Question-shaped headings and direct answers in the opening lines of a page make the difference between being summarised and being skipped.
What they reward: classic local SEO strength, an active Business Profile, and pages structured for extraction.
Perplexity: retrieval-first and citation-hungry
Perplexity searches on every query and attaches citations to nearly every sentence it writes. That makes it the most transparent of the four and the most sensitive to page quality. It favours pages that answer the question high on the page, load fast, and read cleanly as text. Recency counts for more here than on the other engines; stale pages fall out of its citations quickly.
For a niche query like "accountant for medical professionals Sydney", Perplexity often cites a specific service page rather than a homepage. Firms with one page per niche collect those citations; firms with one generic services page do not.
What it rewards: direct answers, fresh content, and a dedicated page for each niche you serve.
Claude: cautious retrieval, verifiable sources
Claude searches the web when a question needs current or local information, and it hedges more than the others when sources conflict. In our testing it leans toward sources it can verify: official registers, established directories, pages whose claims are concrete. Inconsistent details across your web presence hurt more here, because conflicting facts push Claude from naming firms toward generic advice.
What it rewards: consistency and verifiability, including registration details a sceptical reader could check.
What works on all four
- Facts an engine can lift. "Registered tax agents serving hospitality businesses across the Inner West since 2011" works everywhere. Adjectives work nowhere.
- One story across the web. Site, Google Business Profile, and directories agreeing on your name, address, services, and registration.
- Plain, crawlable HTML. Every engine's bots read static text; several stumble on JavaScript-rendered content.
- A page per real question. Niche pages collect citations on every engine, for different reasons.
How to find out where you stand
Ask each engine the questions a real prospect would ask, in Australian terms: "best accountant for tradies in Penrith", "who can lodge my overdue BAS". Ask more than once. Engines vary their answers run to run, which is why our Scorecard averages repeated runs per engine instead of trusting a single response. The pattern across runs tells you which engines already know you, which ones can't verify you, and which competitors own the answers you want.