Four Research Jobs, Not One Subscription
"AI research" hides four separate jobs, and most teams pay two vendors to do the same one. Split them and the choice gets cheaper, easier to defend, and much easier to verify.
Finding sources, reading long documents, synthesising material you already trust, and handling mixed media look identical in a demo. They behave nothing alike once accuracy matters.
The Shortlist, With Status
Each row links to a tracked status page carrying the last automated check, a confidence score, and any death signals we picked up. Never adopt a research tool from its landing page - open the status page and look at the date.
| Tool | Best job | Status | Watch for |
|---|---|---|---|
| Perplexity AI | Citation-first answers | Active | Source quality varies by topic; see Perplexity AI alternatives |
| You.com | Customisable AI search | Active | Frequent product pivots; see You.com alternatives |
| ChatGPT | Synthesis from your sources | Active | Confident gaps when browsing is off; see is ChatGPT dead? |
| Claude | Long-document analysis | Active | Won't fetch the web for you; see Claude alternatives |
| Bard / Gemini | Multimodal inputs | Active | Repeated renames and tier reshuffles |
| Neeva | Ad-free AI search | Shut down | Acquired, consumer product closed |
| Galactica | Scientific literature | Shut down | Withdrawn days after launch over fabricated sources |
Cross-check the survivors on the leaderboard, and if you are choosing between the two general assistants, the ChatGPT vs Claude comparison covers the research differences directly.
The 60-Second Citation Check
This is the whole discipline. Four steps, one minute, applied to every claim you intend to publish or present. It is faster than defending a bad number later, and it catches the failure modes that killed Galactica.
For anything consequential, keep the primary source in a reference manager rather than a chat thread. Bodies like the NIH, OECD, and arXiv publish the originals that answer engines paraphrase, and their URLs outlive any vendor.
Shutdown Risk in This Category
Research tools die from a specific combination: a commodity model underneath, a costly index on top, and users who leave the moment a rival answers better. Here is how the risk splits across the shortlist.
Why Research Tools Die
The two headline failures in this category teach different lessons, and both are still repeating.
Neeva built a genuinely good ad-free search engine with AI answers on top. The product was not the problem - distribution was. Consumers do not change search engines for principle, and subscription search could not out-spend default placement. It was acquired, and the consumer product closed.
Galactica failed the other way. Trained on scientific literature, it produced fluent papers with references that did not exist, and was withdrawn within days. The lesson stuck: in research, fluency without verifiability is a liability, not a feature.
Between those poles sits the ordinary risk. Answer engines with thin differentiation get squeezed when a general assistant ships the same feature for free. When that happens the signals are consistent - a stalled changelog, a quietly disappearing pricing tier, support latency creeping up. Our postmortems on where ChatGPT's promises fell short and where Claude's did show what that pattern looks like even inside healthy products.
Seven-Day Research Stack Audit
Run this once and your research stops depending on any single vendor staying alive.
- Inventory. List every research tool in use, including personal accounts, with owner, plan tier, monthly cost, and which of the four jobs it covers.
- Sample. Take ten claims your team published last quarter and run the 60-second check on each. Record how many fail.
- Sources. Move every verified primary source into a reference manager or shared drive you control.
- Training. Switch model training off on every paid tier and note which tiers do not allow it.
- Export. Pull threads, collections, and saved answers out of each tool, then open a random sample to confirm they are readable.
- Shortlist. One tool per job. Cancel the overlap - most teams pay twice for source-finding.
- Write it down. One page: approved tools, verification rule, what may never be cited from an AI answer, next review date.
Five Mistakes Researchers Keep Making
- Treating a citation list as verification. The reference count is not the evidence. The opened link is.
- Citing the summary, not the study. Numbers drift each time they are retold. Go to the original.
- Ignoring dates. A correct 2021 figure presented as current is still a wrong answer.
- Keeping research only in chat threads. Threads are the least portable storage in your stack; see portable alternatives.
- Assuming free tiers exclude training. They usually do not, and uploaded documents are the riskiest thing you feed them.
What This Actually Costs
Consumer research assistants sit around twenty dollars per seat per month, and team tiers roughly double that in exchange for tighter data terms and admin controls. Enterprise research platforms with private indexes cost far more and are bought by legal or R&D leadership rather than individuals.
The real cost is verification time. Budget a minute per published claim and the tool pays for itself. Skip it and you are buying speed on credit, with a correction as the interest payment.
Price the exit too. Exporting a year of threads and saved sources is an afternoon if you have done it before and a fortnight if you have not.
Keep Exploring the Graveyard
Research overlaps with almost every other AI category, and the same portability questions apply. If you are auditing a wider stack, see AI tools for writing blog posts, AI tools for lawyers, AI tools for marketers, and AI tools for transcription.
Want the underlying data? Browse every tracked tool from the homepage, compare survivors on the leaderboard, read the latest postmortems, or report a tool you believe is winding down.
Bottom line: use one tool per job, open every citation, keep primary sources in storage you own, turn training off, and export monthly. Do that and no shutdown can take your research with it.