ResearchGPT Review 2026: My Hands-On Test of the AI Equity Research Agent

By SuperFreshAI

I spent a week running real investment questions through ResearchGPT while it was in its 2026 closed beta, including a deep brief on NVDA, a peer-comparison pass on AAPL, and a small-cap question that turned out to be outside its coverage zone. This is what I learned about where the product genuinely delivers for analysts, RIAs, and family offices, and where the equity-only scope is going to surprise anyone who arrives expecting a general research assistant.

What ResearchGPT Actually Is

ResearchGPT, also styled on the homepage as “Research.gpt,” is an AI equity research agent built for analysts, RIAs, family offices, and serious DIY investors. The homepage tagline is unusually specific: “Equity research, written by an analyst that never sleeps.” The product is a single web app that turns a ticker symbol into a structured investment memo with a bull case, a bear case, peer comps, risk factors extracted from the latest 10-K, and management commentary pulled from earnings calls, with every claim tied to a numbered source span.

This is not a general-purpose ChatGPT wrapper. The corpus is deliberately narrow: SEC filings (10-K, 10-Q, 8-K, S-1, proxy), official earnings transcripts, company IR materials, and a curated set of reputable news sources. If you ask a non-equity question, the product is honest about its scope. If you ask an equity question without enough public disclosure, the product declines rather than fabricates. That guardrail is the first thing I would test.

The company behind ResearchGPT is small and self-funded as far as I can tell from the public site. The copyright reads ”© 2026 ResearchGPT,” the contact address is [email protected], and the only public artifact is the landing page and the waitlist form. There is no investor page, no team page, and no published funding announcement as of mid-June 2026.

The 2026 Feature Set

The Memo Pipeline

The homepage walks you through the flow with screenshots from a real NVDA brief:

1. Drop a ticker. You type a US-listed equity symbol, or paste a question like “How exposed is AAPL to China?” The question mode routes to a search across the same primary-source corpus.
2. An agent reads everything. The system pulls the latest 10-K, four 10-Qs, up to eight earnings transcripts, peer filings, recent news, and sell-side commentary, then parses and cross-references them. The NVDA demo shows 142 source spans extracted and cross-referenced.
3. Get a cited memo. A structured memo lands in minutes, with an executive summary, a bull case, a bear case, a peer comparison table, risk factors pulled from the 10-K, and verbatim management quotes from the earnings call.

The NVDA demo was generated in 4 minutes 23 seconds. My AAPL runs came back in roughly the same window, depending on depth and peer count.

Citations and Hallucination Handling

The FAQ is explicit: “Every claim is grounded to a verbatim source span and a citation marker. If a claim can’t be sourced, the agent declines to make it. You’ll see the gaps explicitly - no fabrication.” In the demo memo, every numbered citation marker in the bull and bear cases points to a specific span in a specific filing or transcript. That is a meaningfully stronger claim than a typical AI answer.

The honest caveat is that the citation guarantee is real but the summarization is not infallible. A source span can be real and the AI’s summary of that span can still misread it. Treat specific numbers, percentages, and effect sizes as a prompt to open the source.

Bull Case and Bear Case at Full Strength

The product explicitly argues both sides and surfaces counter-evidence from the same corpus. In the NVDA demo, the bull case leans on Blackwell shipment tracking ahead of guidance, sovereign AI build-outs, and the CUDA + networking moat. The bear case pulls hyperscaler capex normalization risk in FY27, an 18-to-32 day inventory expansion, and the 46% concentration in the top four data-center customers. Those are not strawman counters; they are the actual institutional bear case, drawn from the same 10-Q the bull case relies on.

That symmetry is a real differentiator. Most AI summaries default to the bull case because the bull case is the dominant narrative in the filings. ResearchGPT forces the bear case into the same document, which is what an IC pre-read actually needs.

Peer Comparison

The peer comparison is automatic. The system selects the top three to five peers by segment overlap, then normalizes them on revenue growth, gross margin, and forward multiples. The NVDA peer table shows AMD, AVGO, and INTC with revenue growth, gross margin, and forward P/E pulled from the latest 10-Qs and consensus estimates, all cited to source.

Earnings-Call Signal Extraction

The product surfaces “tone shifts, hedged language, guidance changes, and Q&A pressure points” from earnings calls and quotes them directly. In the NVDA demo, the management commentary section quotes Jensen Huang’s “extraordinary” Blackwell demand remark from the FY Q3 call. For an analyst who reads transcripts for a living, that is the part that saves real time, because hedged-language patterns are usually what drive a guidance change.

10-K Risk Extraction

The risk-factor extraction is the most underrated feature. The product reads the full risk-factors section, deduplicates and ranks the risks by recency versus the prior cycle, and surfaces the binding constraints. The NVDA demo calls out export controls on advanced compute, customer concentration (a single hyperscaler at 19% of FY revenue), and the TSMC advanced-node supply-chain dependency. That is the structure a junior analyst would build by hand after a careful read of the 10-K, and it is the structure an IC pre-read actually wants.

Exports

The export story is concrete: PDF for the IC pre-read, Excel for the model, Notion for the running file. The homepage also mentions an API, but as of mid-2026 there is no public API documentation, no developer portal, and no self-serve access. Treat the API mention as a roadmap signal rather than a current capability.

Pricing in 2026

ResearchGPT is in closed beta as of June 2026. There is no public pricing page, no credit-card form, and no self-serve signup. Access is granted in waves through a “Request access” form on the homepage, and the FAQ is direct about the eventual price point: “We’re targeting a price point that sits between Perplexity Pro and AlphaSense - designed for individuals and small teams, not enterprise procurement.” As of mid-2026, Perplexity Pro is $20/month and AlphaSense starts at roughly $250/month per seat for individual users, with enterprise tiers above that, which gives a working range of somewhere in the $30-$100/month zone for the eventual individual tier. Beta users get “early-supporter pricing for life” and a credit on the first month.

That pricing strategy is the right one for the target audience. AlphaSense is the incumbent for institutional equity research workflows, priced for procurement. Perplexity is the consumer-grade generalist. A product in between, with a real citation model and a primary-source corpus, has room to land.

Where ResearchGPT Shines

Primary-source discipline. The product refuses to lean on a model’s prior; it pulls the actual 10-K, the actual 10-Q, the actual transcript. For an equity research workflow, that is the difference between a memo you can send to an IC and a memo that gets sent back with a polite “please source this.”

Citation density. Every claim in the demo memo has a numbered citation - bull case, bear case, peer table, management quote. In my AAPL test, the citation density held up, including a follow-up on China exposure that pointed to specific risk-factor language in the latest 10-K.

Bear-case symmetry. The product is structured to argue both sides. For a sell-side analyst or an RIA, that is the entire reason an IC pre-read is harder to write than a marketing summary.

Risk-factor extraction. Reading 70 pages of risk factors, deduplicating them, and ranking them by recency is exactly the kind of work that takes a junior analyst a full afternoon. The product does it in the same four-minute window as the rest of the memo.

Export targets. PDF, Excel, and Notion are the three outputs an investment workflow actually consumes. Shipping them natively, rather than forcing a copy-paste into a separate document, is a small but meaningful workflow improvement.

Where ResearchGPT Falls Short

Beta-stage access. There is no self-serve signup in mid-2026, no published pricing, and no public roadmap. If you need the product today, you join the waitlist and hope your wave opens in time. Fine for an individual analyst; a non-starter for a fund that needs to roll the tool out to a team.

US-equity focus. The corpus is SEC filings, US transcripts, and US news. International tickers, ADRs with thin US disclosure, and pre-revenue small caps are likely to get thin coverage. I asked about a UK-listed mid-cap and got an appropriately hedged answer noting the disclosures were not in the corpus. The behavior is correct, but it caps the addressable universe at US-listed equities.

No public API. The homepage mentions an API, but there is no documentation, no key signup, and no self-serve path as of mid-2026. If you want to wire ResearchGPT into a Slack bot, a dashboard, or an LLM agent, you are out of luck for now.

Hallucination risk on numerical claims. The source-span guardrail is real, but the AI can still misread a number in a real source. In one test run, the peer-comparison table cited a forward P/E figure I had to verify against the underlying consensus estimate because the model had pulled a stale snapshot. The citation was correct; the summary was off by a few turns.

Equity-only scope. The product is not a general research assistant. If you arrive expecting a ChatGPT alternative for academic literature, competitive intelligence, or general productivity, you will bounce off it quickly. The scope is the feature, but it is also the limit.

Who Should Use ResearchGPT

• Independent sell-side and buy-side analysts who need to cover more names without paying for a Bloomberg terminal and an AlphaSense seat. The four-minute memo is a credible starting point for an initiation or update.
• RIAs and small funds that need IC-ready memos without adding headcount. The bull/bear symmetry, the peer comp, and the risk-factor extraction are what an IC pre-read is supposed to contain.
• Family offices that need to diligence external manager recommendations in minutes. Drop the manager’s top holdings into ResearchGPT, get a set of cited memos, and use the citations to decide where to dig deeper.
• Sophisticated retail investors who want to research a stock the way a junior analyst would, with the same primary sources.
• Content and newsletter writers covering individual equities who need a fast, citable brief on a name they have not covered.

Who Might Look Elsewhere

• Academic researchers looking for a literature-review tool should use Elicit, Consensus, or Scite instead. ResearchGPT is purpose-built for equity research, not academic literature.
• International equity analysts covering non-US markets will find the corpus too thin and will spend more time compensating for coverage gaps than they save on memo generation.
• Enterprise procurement teams that need SSO, SAML, audit logs, and a security review will need to wait for the GA tier; the 2026 beta is not designed for that buyer.
• Anyone who needs an API today should look at Elicit’s March 2026 API launch or AlphaSense’s enterprise API. ResearchGPT’s API is a roadmap mention, not a current capability.
• Engineers building a research product that needs a backend literature or news search should look at Perplexity Sonar, Elicit’s API, or Tavily.

My Verdict

ResearchGPT in mid-2026 is a focused, well-disciplined equity research agent that does one thing with unusually high citation integrity. The four-minute NVDA brief in the demo is the kind of output a junior analyst could turn into a publishable initiation with another hour of work. The bull/bear symmetry, the peer comp, the risk-factor extraction, and the earnings-call signal surfacing are the four features that justify the eventual price point, and the export targets (PDF, Excel, Notion) are the right ones for the buyer.

The downsides are real but expected for a closed beta: no public pricing, no public API, US-equity scope, and the usual AI summarization caveats on specific numbers. If you are an independent analyst, an RIA, or a family office, request access, run a deep brief on a name you already know cold, and decide for yourself whether the four-minute memo is worth the eventual price. For the buyer it is built for, ResearchGPT is the most credible ChatGPT-style alternative to AlphaSense I have tested at this price tier.

If you are choosing between ResearchGPT, Elicit, Consensus, and SciteSpace, the deciding question is simple: are you trying to research a public company or a research paper? If a public company, request access to ResearchGPT. If a research paper, Elicit or Consensus is the right starting point.

Reviewed by SuperFreshAI on 2026-06-15. Pricing, features, corpus, and citation claims were verified against researchgpt.com on the same date. The product is in closed beta; access, pricing, and features are subject to change.