Your buyer opens ChatGPT, picks deep research, and asks it to compare the four vendors on their shortlist. Twenty minutes later they have a cited report.
Everything you’ve done to be retrievable might not matter for that report. In ChatGPT deep research the buyer chooses which sources get read.
OpenAI’s documentation is explicit about it. The user selects which websites, files and connected apps the run may use. They can also restrict web searches to trusted sites, so the work stays on authenticated sources.
That inverts the usual optimisation story, and it changes what’s worth doing.
What Deep Research Does Differently
It’s an agentic mode rather than a faster answer. OpenAI describes it as conducting multi-step research on the internet for complex tasks, accomplishing in tens of minutes what would take a person hours of work.
It Plans Before It Searches
ChatGPT proposes a research plan first, and the user can review and change it before anything runs. So the shape of the investigation is agreed before a single source is opened.
That’s a step no ordinary answer has. It means the framing your buyer brings, and any bias in it, is locked in early.
It Runs Long and Can Be Steered
The run takes minutes rather than seconds, and the user can watch it and interrupt to refine the focus, including adjusting which sources it can reach.
A buyer who notices the report leaning on the wrong kind of source can redirect it mid-run. That’s a meaningful difference from a one-shot answer nobody supervises.
It Returns a Cited Report
The output is a structured document with citations or source links so the reader can verify what it says. Every claim carries its provenance, which is the whole point of the mode.
For a vendor, that means inclusion is visible and checkable. You’re either in the citation list or you aren’t, and your buyer can see which.
The Part That Changes Your Strategy
Source control sits with the user, and OpenAI says so in three separate ways.
The Buyer Picks the Sources
The help documentation states that you choose which sources it can use, listing websites, uploaded files and connected apps. That’s a deliberate product decision, not an incidental setting.
A source list the buyer curates is not a ranking you can climb. It’s a shortlist you’re either on or absent from before the research starts.
Searches Can Be Restricted to Trusted Sites
An update dated 10 February 2026 added two things: connecting deep research to apps, and restricting web searches to trusted sites. A run can now be pointed at authenticated, industry-standard sources only.
Think about what that means for a category with an obvious trade publication. A buyer who restricts the run to it has quietly excluded every blog, every roundup and every comparison page that isn’t on that site.
It Can Read Private Data Too
Deep research can pull from apps and data services the user already has access to, including document stores like Google Drive or SharePoint, alongside authenticated industry sources.
So part of the evidence base may be the buyer’s own files: past proposals, analyst notes, internal comparisons. None of that is reachable by anything you publish.
Why This Is Harder Than Ranking
It’s harder because the gate moved from retrieval to selection, and selection happens before retrieval runs.
Ordinary AI answers reward being findable. If a model searches and your page is a good match, you can be pulled in on merit at query time.
Deep research can be told not to look at you at all. Merit inside an excluded source is worth nothing, and no amount of on-page work reaches past a filter the user set.
That’s not a reason to stop the work described in AI search optimization. It’s a reason to be clear about which of the two problems a given tactic solves.
What Gets You Into a Deep Research Report
Three things, and none of them is a page you publish on your own domain.
Be an Industry-Standard Source
When a buyer restricts a run to trusted sites, they pick the publications their industry already treats as canonical. Trade press, analyst firms, standards bodies, the two or three sites everyone in the category reads.
Presence on those is the whole game for this surface. It’s slow, it’s editorial, and it’s the same work that earns brand mentions in AI search generally, aimed at a much shorter list of destinations.
Be Inside the Sources They Would Whitelist
You don’t need to be the whitelisted domain. You need to be discussed on it.
A category roundup on the trade publication your buyer trusts puts you inside their restricted run without you owning anything. That’s a placement problem, not a content problem.
Be the Primary, Not the Summary
Deep research aggregates and synthesises across sources, so it rewards material that adds something rather than restating what three other pages already said.
Original data, a documented method, a benchmark nobody else publishes: those get cited because they’re the only place the fact exists. A summary of the category gets skipped in favour of whatever it summarised.
What Does Not Work
Keyword-shaped pages aimed at the query rather than the question. A deep research run compares vendors on criteria rather than matching a phrase. A page built for the phrase reads thin beside a source with substance.
How Often It Runs
Less often than ordinary chat, and that shapes how much weight to give it.
OpenAI’s April 2025 update set the allowances at 25 queries per month for Plus, Team, Enterprise and Edu users. Pro got 250 and free users 5, with a lighter version taking over once the allowance runs out.
Those numbers move, so treat them as the shape rather than the figure. The shape is what matters: deep research is rationed, so people spend it on decisions that matter.
A buyer does not burn one of twenty-five monthly runs on idle curiosity. They spend it on a shortlist, a build-versus-buy question, or a renewal. The surface is small in volume and heavy in intent.
How Much One Report Reads
Far more than an ordinary answer, which is what makes the curation question matter so much.
OpenAI’s developer documentation describes its deep research models as able to find, analyse and synthesise hundreds of sources. The stated benchmark is a report at the level of a research analyst.
Hundreds of citation slots is a different opportunity from three. A normal answer names two or three brands; a research report can reference dozens of vendors, publications and datasets in one document.
That cuts both ways. The odds of being included somewhere go up, and the odds of being included as a footnote rather than a recommendation go up with them.
Being Cited Is Not the Same as Being Recommended
A report can cite you as evidence for a competitor’s claim, or list you in a table you come last in. Presence in the citation list is the floor, not the outcome.
Read the sentence your citation is attached to, the same discipline that applies when you check brand mentions in ChatGPT generally.
The Three Ways You Get Left Out
Each one fails at a different stage, and the fix is different for each.
Your Domain Was Never Eligible
The run was restricted and your site wasn’t on the list. Nothing on your domain was read, so nothing about your domain could have changed the result.
The fix is placement on the sources that do get chosen. It’s the slowest of the three and the only one that survives the buyer tightening their filter further.
You Were Read but Added Nothing
Your page was eligible and consulted, and it repeated what four other sources already said. A synthesis step drops the redundant source and keeps whichever one it saw first.
The fix is having something of your own to contribute: a number, a method, a result nobody else can supply.
You Were the Summary, Not the Source
Your page cited someone else’s research, so the report followed the trail and cited them instead. Aggregators lose to primaries in a mode built to trace provenance.
If your best page is a summary of other people’s findings, it’s a signpost pointing at your competitors’ evidence.
How It Sits Beside the Other Surfaces
Three surfaces now sit inside the same assistant, and each one reads something different.
| Surface | What it reads | How you get in |
|---|---|---|
| Ordinary answer | Pages it retrieves at query time | Be findable and quotable |
| Deep research | Sources the user approved, at depth | Be on, or written about on, approved sources |
| Shopping results | A product feed you submit | Submit a catalogue |
The feed route is covered in ChatGPT shopping and product feeds, and it shares nothing mechanically with the other two.
Reading down that middle column is the useful exercise. Three surfaces, three different inputs, and only one of them is a page on your own site.
What to Publish If You Want Citing Here
Publish the thing a report would have to cite you for, because there’s nowhere else to get it.
A Number Only You Hold
Anonymised figures from your own client work, a benchmark across your customer base, a rate nobody else measures. A synthesis step keeps whichever source is the origin of a fact.
It needs a stated method and a date attached. A figure with no method behind it reads as a claim, and a claim without provenance is the first thing a research mode drops.
A Documented Process
Write down how you do the work, in enough detail that somebody could follow it. Process documentation gets cited in comparison reports because it answers the question a buyer is weighing.
This is also the material that survives being summarised. A shallow paraphrase of a detailed method sends the reader back to the method.
A Position With Reasoning
Where the sensible answer is contested, say what you think and show the working. Reports comparing options need the arguments on each side, and a clearly reasoned position is quotable in a way a neutral overview is not.
None of these is a keyword page, and that’s the point. This surface reads like an analyst, so it rewards what an analyst would cite.
How to Test Your Own Position
You test it by running the report your buyer would run, then reading the citation list rather than the prose.
Run It Unrestricted First
Ask for a comparison of vendors in your category with no source restrictions, and note every domain the report cites. That list is your realistic universe.
Then Run It Restricted
Repeat it restricted to the two or three publications your industry treats as authoritative. If you vanish between the first run and the second, you’ve found your gap exactly.
The difference between those two citation lists is the most useful diagnostic this surface offers. It costs two runs.
Read Who Got Cited Instead
Where a competitor appears and you don’t, open the citing page. The question is never why the model liked them, it’s which publication carried them and not you.
What This Means for B2B Buying
The surface matters more than its volume suggests, because of who uses it and when.
A twenty-minute cited report is what somebody produces to justify a decision to other people. It gets pasted into a document, forwarded to a manager, and referenced in a meeting you’re not in.
Being absent from it isn’t a lost impression. It’s absence from the artefact that frames the decision, which is why AI visibility for B2B SaaS treats source presence as the objective rather than traffic.
Frequently Asked Questions
Can I optimise my site to appear in deep research?
Only partly. Your site can be read when a run is unrestricted, but a buyer who limits the run to trusted publications excludes your domain however good it is. Presence on those publications is the real lever.
Does deep research use the same index as normal ChatGPT answers?
It searches the web and can also read files and connected apps the user grants it, which ordinary chat does not. Treat it as a different mode with a wider and more controllable evidence base rather than a longer version of the same thing.
How much is it used?
OpenAI publishes allowances rather than usage. The April 2025 figures were 25 runs a month on Plus, Team, Enterprise and Edu, 250 on Pro and 5 on free. Rationing means the runs that happen are deliberate.
Do citations in a deep research report send traffic?
Some, because the report carries source links the reader can follow to verify a claim. Treat the citation as the objective in its own right, since it reaches people who never click.
Is this the same as agent mode?
They’re related but distinct. A July 2025 update gave deep research access to a visual browser through agent mode. The original deep research option stayed in the tools menu, so the two are selected differently.
What should I do first?
Run the comparison your buyers would run, twice, once unrestricted and once limited to your industry’s canonical publications. The gap between the two citation lists tells you where to aim before you spend anything.
Run the Report Your Buyer Would Run
Open deep research and ask for the vendor comparison a buyer in your category would ask for. Read the citation list at the bottom rather than the summary at the top.
Every domain in that list is a placement target, ranked for you by the thing your buyers are using to decide. That’s a better prospect list than any tool will sell you, and it costs one of your twenty-five runs.