How ChatGPT, Perplexity, Copilot and Gemini choose what to cite

If you’ve watched the same prompt return four different sets of citations across ChatGPT, Perplexity, Microsoft Copilot and Gemini, you’ve already seen the central problem. The major LLMs do not pick sources the same way, and even within a single product the behaviour shifts month to month.

We’ve audited dozens of LLM citation patterns for clients across MSPs, SaaS and ERP consultancies through 2025 and into 2026. Here is what we believe holds up to scrutiny, what’s still guesswork and how we’d advise a B2B tech marketing team to act on what’s known.

The basic mechanics

Every retrieval-augmented chat assistant works in roughly three stages. It interprets the user’s question, fetches a set of candidate sources from an index and then generates an answer that may quote, paraphrase or cite some of those sources. The differences between the products show up at each of those stages.

ChatGPT search uses a mix of Bing’s index and OpenAI’s own crawl. Perplexity uses Bing plus its own crawler and a heavy reranking layer. Microsoft Copilot leans on Bing directly. Gemini uses Google’s index, which makes it close kin to Google AI Overviews. None of these products publish their full grounding stack, but their behaviour has been studied enough that we can describe the patterns.

What gets retrieved

Retrieval is the unglamorous half of the problem and it’s where classic SEO still earns its keep. If your page is not in the underlying index, no clever prompt-shaped writing will get it cited. Crawlable, fast, well-linked pages still win here. We covered the technical side in our technical SEO audit checklist for tech sites.

Within the retrieved candidates, three signals appear to matter most:

• Topical match at the passage level. The model is matching the question to chunks of your content, not your page as a whole. A page that buries the answer under five paragraphs of preamble is at a disadvantage.
• Authority cues at the domain and entity level. Pages from sites that get mentioned alongside the topic in trustworthy places do better. We unpack this in brand mentions vs backlinks in AI search.
• Freshness for time-sensitive questions. Compliance, pricing and product-comparison queries pull more recent content than evergreen definitional queries.

What gets cited from the retrieved set

Retrieval is not citation. The model fetches more than it shows. Whether your source ends up named in the final answer depends on how usable it is.

In our testing, sources that get cited share three traits. They state the answer plainly. They have clear authorship or institutional voice. They are easy to quote without distortion. The third one is underrated. If your sentence has to be rewritten to make sense out of context, you’ve made the model’s job harder.

We’ve written more about how to draft for this in writing content that AI search engines actually cite. The short version is that conversational, hedged prose loses to direct, declarative prose for citation purposes.

How the four products differ in practice

A practical comparison from our testing through Q1 2026:

These behaviours are not stable. Perplexity shifted markedly between September 2025 and March 2026. ChatGPT search’s citation density changed twice in the same window. Treat any single audit as a snapshot, not a rule. We’ve drawn out the differences between two of the most-asked-about engines in Bing Chat vs ChatGPT citations.

What this means for B2B tech sites

A few takeaways we’d stand behind for an MSP, SaaS firm or IT consultancy trying to get cited:

• Lead with the answer. The first 150 words of any post matter disproportionately. State the claim, then justify it. Burying the lede is a citation-killer.
• Name your entities. If a model needs to know what your product does, what it integrates with and who it’s for, those facts should sit on your site as plain prose, not buried in a feature matrix.
• Build a third-party footprint. Trade press, podcasts, analyst mentions, comparison sites and increasingly Reddit threads in AI search results. LLMs use the broader web to decide who counts as authoritative on a topic. Backlinks still help. Mentions without links also help, more than they did for classic SEO.
• Add schema where it earns its keep. Article, FAQPage, Product, Organization, Person. We’ve gone deep on this in structured data for AI search and the related piece on schema markup for SaaS websites.
• Audit your visibility regularly. Citation patterns drift. Our piece on auditing your visibility in Copilot and ChatGPT lays out a repeatable process.

What we still cannot prove

We will not pretend the picture is clean. Several questions remain genuinely open:

• How heavily each model weights traditional backlink signals versus brand mentions in unlinked text.
• Whether llms.txt influences any of the major engines as a citation signal rather than a discovery hint.
• How long a citation pattern persists once established. Some clients see a page cited for months. Others lose visibility after a single product update.
• Whether paywalled content gets retrieved at all, and how partial-paywall snippets behave.

We test these things on real client sites and on our own properties. The honest answer for most of them is “it depends, and the dependencies are not stable”.

What we’d do this quarter

If you have one quarter and a constrained budget, we’d suggest the following sequence. Pick fifteen buyer-intent prompts that matter to your business. Run them across the four products. Note which competitors get cited, what shape of content the citations point to and where you are absent. Pick the three biggest gaps and write or rework pages to fit the answer shape. Re-run the prompts in six weeks. You will not get clean before-and-after data, but you will learn faster than you would by reading think-pieces.

That kind of structured experimentation sits inside the broader work we do as part of our AI SEO services. It rewards teams that are willing to publish, measure, learn and iterate without waiting for the discipline to settle.

Working through any of this on your own site? We’d be glad to compare notes.