AI Presentations with Real Citations: What “Sources” Should Actually Mean

By the Kinsy team · Published 25 July 2026

An AI deck maker that invents a statistic hasn't got a bug — it has handed you a career problem. If you present a hallucinated number to a client, a board or an examiner, the tool doesn't take the question afterwards. You do. This is what genuine citation support in an AI presentation should look like, what the leading tools actually do about it, and how to audit any deck before you stand up.

Full disclosure: we build Kinsy, and sourcing is one of the things we build it around. The criteria in this post apply to us exactly as much as to anyone else — hold every tool, ours included, to them.

Can AI presentations include real citations?

Yes — a small number of tools now retrieve live sources while they generate and attach numbered citations you can actually check; the rest write from model memory, and model memory is where hallucinations come from. The distinction matters more than any feature list. A tool that generates a deck and then decorates it with plausible-looking references has not cited anything; it has produced two hallucinations that agree with each other. A real citation is a claim that exists because a retrieved source supports it, with a visible trail from the bullet on the slide to a document you can open.

Why decks are where hallucinations hide best

A chatbot answer at least keeps its hedges — the “reportedly”, the “according to some estimates”. A slide strips them out by design. Deck writing compresses every claim into its punchiest form: a paragraph of qualified reasoning becomes a five-word bullet with a bold number, because that's what good slides do. The compression that makes a deck persuasive is precisely the step where provenance disappears. Nobody footnotes a bullet by hand, the qualifier was the first thing cut, and by the third revision nobody remembers which numbers came from the source document and which the generator supplied on its own.

Then the deck travels. It gets forwarded, presented, excerpted into other decks — each hop further from anyone who could say where a figure came from. And unlike a chat answer you skim once, a deck is a performance with your name on the title slide. In consulting, a fabricated market size in front of a client is a fired-agency problem. In a boardroom, an invented growth figure can end up in minutes and filings. In a viva or a thesis defence, an unsupported claim is the exact thing the examiner exists to find. The slide format hides hallucinations better than any other medium — which is why sourcing has to be built in at generation, not patched on before the meeting.

What genuine citation support looks like

Treat these four as pass/fail criteria for any AI presentation maker that says it does “sources”:

• Live retrieval at generation time. The tool searches while it writes, and claims are grounded in what it just read — not reconstructed from whatever its model absorbed in training. Anything generated first and referenced later fails.
• Numbered per-claim citations on the slide. Not a vague “sources consulted” appendix — a marker on the specific bullet, pointing at the specific document that supports it. If you can't tell which source backs which claim, you can't check either.
• A checkable source list. Real, resolvable references you can open before the meeting. A citation you cannot follow is decoration.
• A fact-check pass before you present. Someone — the tool, then you — re-reads every claim against its source before the deck is called finished. Generation and verification are different jobs; a tool that only does the first has done half the work.

What the leading tools actually do about sources

The field is moving, so treat everything here as a snapshot — as of July 2026 — and re-verify before you rely on it.

Gamma made real progress here: its Agent, the headline feature of the Gamma 3.0 release in September 2025, can research the web on request and pull citations in to support key points. That's genuinely better than generate-from-memory. From Gamma's public material we couldn't confirm that every factual claim gets a per-claim, on-slide citation by default — coverage appears to depend on what you ask the Agent to do — so audit the output rather than assume it.

ChatGPT can browse, and when its search mode is on it returns linked sources in the conversation; since May 2026 an official add-in will also build and edit slides directly inside PowerPoint. But the sources live in the chat, not in the file — the deck you export doesn't carry the provenance with it. And when it isn't browsing, its well-documented failure mode is fabricating reference details — author names, identifiers — that look exactly like real ones, which is worse than no citation at all.

Perplexity is retrieval-first by architecture: it searches on every query and attaches numbered citations to its answers as standard, which is why researchers like it. Its Labs tier (on Pro, $20/mo) can carry that research into slide-style exports. The deck side is younger than the answer engine, though, and how much of the citation trail survives into a file you'd actually present varies by project — check yours before trusting it.

Vendor capabilities as of July 2026 — all of these products change quickly, so check the vendors' pages before deciding. And the disclosure above applies doubly to this table: we build Kinsy, so verify every row yourself.

How Kinsy handles sources

Kinsy treats sourcing as part of generation, not an afterthought. It plans the narrative first, and when a topic needs outside facts it runs live research while it writes, putting numbered citations on the slides themselves with a source list you can open and check. Before you ever see the deck, it fact-checks its own claims — the verification step is a separate pass, not a promise that generation got it right the first time. Every slide ships with speaker notes, which is where the context behind a cited figure belongs: the slide stays punchy, the backing travels with the file.

One more detail matters for citations specifically: the deck exports as a native, fully editable .pptx, so citation markers and source lines are real text you can edit, restyle or delete — not pixels baked into an image. Provenance that arrives flattened is provenance you can't maintain; we've written elsewhere about why AI deck tools break on export and why everything staying an editable object is the bar.

How to audit any AI deck's claims in 10 minutes

Whichever tool made the deck — ours included — run this before you present it:

1. Mark every claim-shaped thing (2 min). Go slide by slide and mark each number, percentage, date, named organisation, quotation and superlative — “largest”, “fastest-growing”, “first”. These are the statements someone in the room can challenge.
2. Demand an origin for each (2 min). For every mark, ask: where did this come from? Your brief, a cited source, or nowhere? Anything with no visible origin is unverified by definition, no matter how plausible it reads.
3. Open the citations that do exist (3 min). Follow each numbered source and confirm it actually says what the slide says. A real link that doesn't support the claim is the subtlest failure — citation is not the same thing as support.
4. Check the dates (1 min). A true figure from 2021 presented as the current state of the market is its own kind of wrong. Confirm the freshest claims are actually fresh.
5. Soften or delete the orphans (1 min). Any claim you couldn't trace either gets a hedge you're comfortable defending, gets replaced with a number you can stand behind, or comes off the slide.
6. Move the defence into the notes (1 min). Put the source names and context for your most challengeable claims into the speaker notes, so the answer to “where's that from?” is already under the slide when someone asks.

Ten minutes. It's the difference between presenting a deck and vouching for one.

Keep reading

• How Kinsy researches and cites sources in your decks
• Turning a research paper into a presentation without losing rigour
• AI presentation makers that export truly editable PowerPoint files