Can You Control What Sources Perplexity Uses? I Ran 75 Searches to Test It

If you run a brand, you have probably wished you could tell an AI engine which sources to trust.

Cite our official pages. Use real reviews, not random listicles. Look at what users actually say.

I spent an afternoon running 75 searches on Perplexity to find out whether you can.

The short answer: barely. You can nudge it. You cannot really steer it.

The longer answer was more useful. The way it failed told me exactly where a brand's evidence was strong, where it was missing, and where it was quietly being written by someone else.

Key findings#

• Of the four source instructions I tested, only "use Reddit" worked in all five queries. Reddit threads exist for almost any tool question and Perplexity does not lean on them by default, so there was always something to pull and room to add it.
• "Use official pages" worked only on brand-specific questions. On broad "best tool" questions, the brands' own pages were barely in the web Perplexity searched, so there was nothing to pull.
• "Use reviews" was the messiest. Of 116 review articles Perplexity cited, 50 were written by vendors, the tool companies themselves. The most-cited "review" I found was a direct competitor's.
• "Avoid listicles" mostly just made answers thinner. Baseline answers cited about 13.5 sources on average; "avoid listicles" answers averaged 9.9. Perplexity did not replace listicles with better sources, it just had less to say.
• A source instruction is a flashlight, not a remote control. It reveals which evidence layers about your brand exist and have room, not what Perplexity will choose to trust.

What I tested#

I took one category, AI meeting note tools, and asked Perplexity the same five kinds of questions under different source instructions: a normal baseline, then "use official pages," "use reviews," "use Reddit," and "avoid listicles."

I ran 75 searches in total: five queries, five instructions, three repeats each.

The repeats mattered. Perplexity does not return the same sources every time. The mix drifts from run to run. So I did not count an instruction as "working" unless it moved the answer beyond the normal drift I already saw in the baseline runs.

Then I looked at every source it cited, all 926 of them, and sorted each one by what it actually was: an official page, a review platform, a Reddit thread, a YouTube video, a listicle, a comparison page, and so on.

Here is what happened.

Reddit was the only instruction that worked every time#

When I asked Perplexity to use Reddit or user discussions, it listened in all five queries. This was the cleanest result in the whole test.

The reason is boring, but it matters for everything else here. Reddit threads exist for almost any software question, and Perplexity does not lean on them by default. So when you ask for them, there is always something to pull, and room to add it.

Those two things, something to pull and room to add it, decided every other result in the test.

Official pages worked, but only when there was room#

"Use official pages" worked on some brand-specific questions. On a broad question like "best AI meeting note taker for sales calls," it barely moved anything.

The reason was not that Perplexity ignored me. On a question like "best tools," the brands' own pages were almost absent from the web Perplexity was searching. There was nothing to pull. And on some brand-vs-brand questions, the brands' official comparison pages were already in the baseline, cited whether I asked for them or not. There was no room to add more.

Nothing to pull, or no room to add. Same instruction, both kinds of failure.

"Use reviews" was messier than it sounds#

This was the surprise.

"Review source" sounds clean. You picture G2, Capterra, a careful hands-on writeup. But Perplexity's review layer was a tangle of review platforms, YouTube videos, blog posts, comparison pages, and vendors reviewing each other.

Out of 116 written review articles Perplexity cited across the test, 50 were written by vendors, the tool companies themselves. The most-cited "review" I came across was tl;dv's "honest review" of Fathom. tl;dv competes with Fathom directly.

So "use reviews" never produced a clean result, and it failed for two opposite reasons. On questions like "Fathom review," the answer was already almost entirely reviews. Nothing left to add. On broad questions, there were barely any independent reviews to reach for. The layer was either full or empty. Neither had much to do with how I phrased the request.

"Avoid listicles" sometimes just made the answer thinner#

When I asked Perplexity to avoid listicles and roundups, it could do it on the broad "best tools" question, where listicles dominate. But something else happened: the answer got thinner.

Baseline answers cited about 13.5 sources on average. "Avoid listicles" answers averaged 9.9. When I took the listicles away, Perplexity did not replace them with better sources. It just had less to say.

So, can you control what sources Perplexity uses?#

Sort of. But that turned out to be the wrong question.

A source instruction is not a remote control. It is more like a flashlight.

It does not let you decide what Perplexity will trust. It shows you which evidence layers are strong enough to be found. When an instruction works, that layer is there and has room. When it does nothing, that layer is either missing, or already full of content someone else wrote about you.

That changed how I think about AI visibility. I would no longer only ask, "Does my brand show up?" I would ask:

• Which source layer is actually carrying the answer?
• Which layer is missing entirely?
• Which layer is partly being written by competitors?
• Which layer gets thinner the moment you remove listicles?

Those questions point at real work. If competitors are writing part of your review layer, that is a gap you can close. If a whole layer goes empty when you ask for it, you have nothing there to be found. The instructions fix none of this. They just show you where you stand.

A few honest limits#

This was one engine (Perplexity), one category (AI meeting note tools), and 75 searches over a short window. Perplexity is also one of the friendlier engines for source-aware search, so if anything, this is the optimistic case. Other engines are likely to listen less, not more. I would not read these exact numbers as universal.

But the shape of it has held up everywhere I have looked since: you can nudge sources, but you cannot summon ones that are not there.

You can explore every cited source yourself: filter, sort, and download the full source data, all 926 citations, each traceable to the exact search and answer it came from, plus rollups by URL and by domain.

FAQ#

Can you force Perplexity to cite specific sources?#

No. You can nudge it toward a source type, but you cannot force it. In this test an instruction only worked when that kind of source already existed for the query and the answer had room to include more. When the source layer was missing or already full, the instruction did nothing.

Which source instruction worked most reliably?#

"Use Reddit." It moved the answer in all five queries, because Reddit discussions exist for almost any tool question and Perplexity does not cite them by default, so there was always something to pull and room to add it.

Why did "use reviews" not give clean results?#

Because the review layer was either already full (on "X review" questions the answer was mostly reviews already) or nearly empty (on broad questions there were few independent reviews). On top of that, of 116 review articles cited, 50 were written by vendors, so "reviews" often meant competitors reviewing each other.

Does this apply to ChatGPT or Google AI Overviews?#

This tested only Perplexity, which is one of the friendlier engines for source-aware search. Other engines are likely to follow source instructions less, not more, so treat this as the optimistic case rather than a universal rule.