Some investing questions are about a particular day. You bought Caterpillar on the 16th and want to know what the company had reported by then. A post says Delta's revenue is down 68% and you want to know whether that was true when it was written. Your mother's income plan needs Procter & Gamble's latest declared dividend, and the dates, as they stood last Friday. The answer has to come from what the company or the fund itself had published by that day, not from whatever is on the web this morning.
AI assistants will now go and find those documents for you. We gave ChatGPT dots, Grok Bot, Instinct and Meta's Muse the same 18 questions about real US companies and funds, each asking for the facts as they stood at 4 pm Eastern on Friday 16 Oct 2026, and asked them 10 days later, when newer pages were already online.
Let it find a company's or fund's own reports and pull out the numbers, which is fine for a first look, but check the report before you act on a figure. It may give you last quarter's results when newer ones are out (Muse did), use a fund's holdings from a different date than the one you asked about (Instinct did), or quote a figure that isn't in the report it links (Grok Bot did). So open the report it links and check that it covers the period you mean and really has the number. Start with ChatGPT dots on the paid Pro plan, which avoided these mistakes, though it sometimes skipped part of a question, so check it answered everything.
In our first investing review, the same assistants answered 54 questions from documents we put into the chat with them. ChatGPT dots leads both. Read the two reviews as two different jobs. The first review scores the research itself; this one scores whether an assistant can find its own evidence and stay within the day you give it.
Overall ranking · researching real stocks and funds, score out of 100
dots leads; Instinct and Muse each have two answers scored 0, Grok Bot one
Solid dot: the total, out of 100. Hollow marks: three of its four parts, scored yes or no on every answer.
TotalCorrectTrustworthyUsable
- Total = 40 Correct + 30 Trustworthy + 20 Usable + 10 Second try, averaged over six equal jobs. Second try is 88.3–92.3 for every assistant and is not drawn.
- Totals count an answer with a critical error or no answer as 0; the part marks average the answers given.
- The axis starts at 60.
- Plans differ: ChatGPT Pro, X Premium, Meta’s free plan for Muse; Instinct shows no plan.
- Red marks the best result on each measure; ties are marked equally.
Can it read a company's latest results?
Mostly yes: ChatGPT dots, Grok Bot and Instinct read Johnson & Johnson's newest quarter, and Muse used the one before. Asked whether Johnson & Johnson had really grown in its latest quarter, given "two different earnings numbers", the answer had to come from the quarterly release the company filed on 13 Oct 2026, three days before the date. Sales grew 5.9%, to $25.4 billion, while reported earnings per share fell to $0.09 from $2.12, mainly because of a litigation charge; the adjusted figure was $2.59. ChatGPT dots and Instinct gave all of that, with the reconciling items, and met every check. Grok Bot gave the same figures and missed one of the other checks. Muse answered from the second-quarter release, published in July, when the third-quarter one was out before the date, and corrected itself when asked to check again. Its first answer met 2 of the 6 correctness checks.
ChatGPT dots scored 97.6 on company research, Instinct 96.7, Grok Bot 94.3 and Muse 77.8. Asked what Caterpillar does, where its revenue comes from and what risks it reports itself, each assistant worked from Caterpillar's 2025 annual report. Asked about conflicting reports that Starbucks wanted to buy Chipotle, each assistant kept the newspaper reports apart from anything either company had confirmed, and treated the takeover as unconfirmed.
Can it check a claim you read online?
Yes: every assistant caught a misleading post and checked two other claims against the companies' own filings. A post said Delta Air Lines' revenue was down 68%, because it brought in $20.2 billion last quarter against $63.4 billion in 2025. Every assistant said the post compared one quarter with a whole year, and worked out from Delta's own figures that revenue was up about 21% on the same quarter a year earlier. Asked whether Robinhood Markets "doubled its revenue last year", each assistant found the 2025 annual report, gave the two years' revenue and said that growth of about 52% is not doubling. Asked whether Home Depot's chief executive had sold half of their shares last month, each assistant checked the insider filings up to the date and found none showing such a sale. Grok Bot and Instinct met every check; ChatGPT dots met 4 of its 6 correctness checks. Muse's second try on this question arrived after our limit, so it counts as none.
Instinct scored 96.7 on checking a claim, ChatGPT dots 94.4, Grok Bot 93.3 and Muse 87.8.
Can it tell you what a fund held on that day?
Not reliably: ChatGPT dots stayed within the date and said what could not be known about that day's holdings, and Instinct twice worked from holdings lists dated a week later. Three questions asked about two iShares funds, the Core S&P 500 ETF and the Russell 1000 Growth ETF, for a fictional investor who holds both: his exposure to the largest holding the two share, how much they overlap, and whether their published holdings are current. The funds' holdings files are updated often, so the file an assistant finds 10 days later is dated after the date we asked about, and the question itself said what to do then: name the file and its date, say what cannot be established as of the date, and show the method.
Grok Bot did that on two of the three questions and met every check on both. ChatGPT dots stayed within the date on each of them, scoring 92.0, 100.0 and 90.0, and missed one correctness check on two. Instinct worked both the exposure and the currency questions from holdings files dated 23 Oct, a week after the date. It said the files were later than the date, but used their weights to work out the exposure in one answer and their dates to work out how far the holdings run behind in the other. Both answers score 0 under our rule on pages from after the date. Muse gave no answer to the exposure question within the 15 minutes we allow; a full reply came later and does not count, so that answer scores 0 too. On the overlap question, Grok Bot's answer cited one fund's filing for a holding weight the filing does not show: Amcor at 2.84% of the S&P 500 fund, where the filing gives 0.031%. It called the line implausible in the same answer and corrected it when asked to check again, but citing a filing for a figure it does not contain counts as a critical error, and that answer scores 0.
ChatGPT dots scored 94.0 on fund holdings, Grok Bot 66.7, Muse 50.0 and Instinct 25.0. Three questions decide a job score, so one answer scored 0 moves it by a third.
Can it update your view when results land?
Mostly: ChatGPT dots, Grok Bot and Muse checked an investor's conditions against PNC's own materials, and Instinct missed the guidance. A fictional investor had written a thesis on PNC Financial Services after its second-quarter results: revenue growing at least 12%, net interest margin at or above 2.90%, and no cut to guidance. The third-quarter release showed revenue up 10.9%, below his line, and margin at 2.94%, above it. PNC's presentations showed full-year revenue guidance raised, not cut. ChatGPT dots and Muse checked the conditions against those materials and met every check. Grok Bot reached the same conclusions and missed one other check. Instinct met 3 of 6 correctness checks: it said the guidance condition could not be settled from the releases, but had not read the presentations filed before the date.
Asked whether UnitedHealth Group's guidance changed after headlines said it "beat expectations", each assistant gave the July and October guidance as the company stated them and said that a "beat" refers to expectations the release itself does not state. Asked which version of AZZ's amended quarterly filing counts, each assistant found that the amendment, filed the day before the date, added only a share-repurchase table, so the original's financial statements still stand. ChatGPT dots and Muse scored 100.0 on this job, Instinct 91.7 and Grok Bot 90.0.
Can it confirm a dividend from the source?
Yes for a declared dividend: each assistant found Procter & Gamble's declaration and its dates. Asked for Procter & Gamble's latest declared dividend and its dates, each assistant found the declaration of 13 Oct 2026: $1.0885 a share, record date 23 Oct, payable from 16 Nov. ChatGPT dots met every check; Grok Bot and Instinct gave the same facts at greater length than the question needed, and Muse did that and also left a limit unstated.
Only Muse gave the JPMorgan fund's latest payment from the fund's own record, and only Instinct met every check on the Costco question. Asked what JPMorgan's Equity Premium Income ETF pays and how often, only Muse gave the latest payment from the fund's own distribution record, $0.34134 a share with a 1 Oct ex-date, and said that a yield it could not show was published by the date is not a fact as of that date. It also said it could not show that the distribution table or the fact sheet was online by the date; the October payment came from the table as it read it 10 days later. ChatGPT dots met 1 of 4 correctness checks here and Instinct 2 of 4. Asked whether Costco was about to pay a special dividend, each assistant said that no special dividend had been declared and gave the regular 14 Oct declaration; Instinct also listed the January, April and July declarations and was the only assistant to meet every check on this question.
Instinct scored 81.7 on the three dividend questions, Muse 80.7, Grok Bot 79.2 and ChatGPT dots 77.5, its lowest job.
Can it tell you what happened and what is next?
ChatGPT dots got every date right, and Muse took Nike's date from a page updated after the day. Asked when Kimberly-Clark, Netflix and Nike would next report, and which dates the companies had confirmed, the right answer as of 16 Oct was two confirmed dates, Kimberly-Clark on 5 Nov and Netflix on 20 Oct, and no confirmation found for Nike. ChatGPT dots gave exactly that and met every check. Muse gave a 17 December date for Nike from a calendar page updated on 23 Oct, a week after the date. It called the date an estimate rather than a confirmed one, and withdrew it when asked to check again, but an answer that uses a page from after the date scores 0 whether or not it labels it. Grok Bot said Nike had not announced a date, which its sources could not show, and met 3 of 5 correctness checks.
Asked what Kimberly-Clark had announced in the week before the date, each assistant listed the company's own filings from that week, dated: the early results of its exchange offers on 13 Oct and its early third-quarter figures on 15 Oct. Asked when Nike's next results would be, when the company had not announced a date, each assistant gave Nike's recent reporting pattern as a pattern, not a date. Instinct said Nike reports on Thursdays, which one of the dates it listed contradicts, and missed that check.
ChatGPT dots scored 100.0 on news and what is next, Instinct 90.7, Grok Bot 86.7 and Muse 63.3.
Score by job · researching real stocks and funds
Fund holdings at the date were the hardest job for Grok Bot, Instinct and Muse
Number: the job score out of 100. One square per first answer: filled if it met every check with no critical error.
Met every checkMissed at least one check, or no answer
| Assistant | Company research3 questions | Checking a claim3 questions | Holdings and records3 questions | Investment view3 questions | Dividend income3 questions | News and decisions3 questions |
|---|---|---|---|---|---|---|
| dots | 97.6Best | 94.4 | 94.0Best | 100.0Best | 77.5 | 100.0Best |
| Grok Bot | 94.3 | 93.3 | 66.7 | 90.0 | 79.2 | 86.7 |
| Instinct | 96.7 | 96.7Best | 25.0 | 91.7 | 81.7Best | 90.7 |
| Muse | 77.8 | 87.8 | 50.0 | 100.0Best | 80.7 | 63.3 |
- Listed alphabetically, not ranked. The checks: the question’s own yes-or-no questions, sources dated by the cutoff, the right period, honesty about limits, answer first and concise.
- Red marks the best result on each measure; ties are marked equally.
| Job (3 questions each) | ChatGPT dots | Grok Bot | Instinct | Muse |
|---|---|---|---|---|
| Company research | 97.6 | 94.3 | 96.7 | 77.8 |
| Checking a claim | 94.4 | 93.3 | 96.7 | 87.8 |
| Holdings and records | 94.0 | 66.7 | 25.0 | 50.0 |
| Investment view | 100.0 | 90.0 | 91.7 | 100.0 |
| Dividend income | 77.5 | 79.2 | 81.7 | 80.7 |
| News and decisions | 100.0 | 86.7 | 90.7 | 63.3 |
Cells show each assistant's average score out of 100 on that job's three questions.
Does it get to the point?
ChatGPT dots did every time, and Grok Bot's answers often ran long. Two checks measure whether an answer is easy to use: whether its first two sentences give the answer, and whether it is concise, with nothing a third of its length could lose without losing what was asked. ChatGPT dots led with the answer and was concise in all 18 answers. Grok Bot led with the answer in all 18 but was concise in 8, often adding context the question did not ask for. Instinct led with the answer in 17 of 18 and was concise in 14. Muse led with the answer in all 17 it gave and was concise in 12.
Does asking it to check again help?
Sometimes: it fixed two of the four serious mistakes, but not Instinct's two answers built on later holdings files. After each first answer we sent the same request: recheck against the original question and the sources, and correct anything unsupported. Asked to check again, Grok Bot corrected the Amcor figure, and Muse withdrew the Nike date and, for a lesser slip that was not one of the serious ones, switched to the right Johnson & Johnson quarter. The second try counts for 10 of the 100 points; the first answer carries the rest, and a critical error scores the whole answer 0, even when the second try corrects it.
Which review should you read?
Read the first review if your question comes with its documents, and this one if it is about a particular day. Our first investing review, published on 23 Oct 2026, gave the same assistants 54 questions with the documents already in the chat: results, fact sheets, statements and posts. That test measured the research. This one gave 18 questions with nothing attached and measured whether an assistant can find the company's or fund's own document and stay within the day you ask about. The scoring is the same, and ChatGPT dots leads both. Here Grok Bot, Instinct and Muse are each within about one answer of the next. The answers scored 0 also differ: in the first review one answer scored 0, an answer Instinct never gave; here five did: four for critical errors—three uses of later information and one figure its cited filing does not show—and one for a missing answer. The two scores are not comparable and we do not combine them.
So, can you rely on AI to research a stock or fund?
Hand it the legwork of finding earnings reports, filings and dividend notices, and before you act on a figure, check that the report it links covers the period you mean and really contains the number. Start with ChatGPT dots on the paid Pro plan, which never gave a figure from the wrong period, and check that it answered everything you asked.
To start, ChatGPT dots, on ChatGPT Pro (a recorded USD 200 a month; cheaper ChatGPT plans were not part of this review), needed the least checking: it led with the answer, kept it short and used nothing from after the date, though it was last on dividends, where it held back a fund's latest payment it could not date. Grok Bot got the most facts right but makes you read through long answers, and one of its answers cited a filing for a figure the filing does not contain. Instinct read the companies' own results as well as the leader and was the only one to give everything the Costco dividend question asked for, but twice it used funds' holdings files from a week later, though it said so. Muse, on Meta's free plan, was the only one to give the fund's October payment from its own record, while saying it could not show that record was online by the date, but it answered from an old quarter once, from a later calendar page once, and not within the time limit once.
Before you act, glance at the date on the document an answer links: that is where Instinct's two serious mistakes and Muse's showed, and those answers said so themselves; Grok Bot's was a figure the cited filing does not contain, which the answer itself called implausible. Muse's use of an older quarter's report was a separate, lesser slip. Asking an assistant once to check its answer fixed Grok Bot's figure and Muse's calendar date, but not Instinct's two. Nothing here is investment advice.
Method and where the evidence stops
Each of the 18 questions came from one of the same four fictional investors as our first investing review and asked for a link to the company's or fund's own document, with its date. They cover the six jobs of our first review, three each: reading a company's results, checking a claim you read somewhere, reading what a fund holds, updating your view when results land, confirming a dividend, and sorting out what happened and what is next. Each assistant ran in its own product, with the same account and conversation setup as our first investing review: ChatGPT dots on ChatGPT Pro with memory off, Grok Bot on X Premium, Instinct over WhatsApp with no plan shown, and Muse on Meta's free plan, opening a new chat for each question. The conversations of the first review continued into this one and could not be reset, so the answers are not independent. The questions were fixed on 21 Oct 2026, before any assistant was asked, and each has a key of the facts the answer must contain, taken from the companies' and funds' own documents published by the date, and the things it must say cannot be known as of that date. Each question ran once.
The questions went to every assistant between the evening of 26 Oct and the early morning of 27 Oct, Pacific time, within 181.3 seconds of one another, with 15 exceptions. Our sessions running ChatGPT dots and Instinct stalled overnight, so 13 of their questions, 8 for ChatGPT dots and 5 for Instinct, were sent 13 to 32 hours after the same question went to the others. Grok Bot was asked one question 35 hours later and Muse one 33 hours later, after we missed their time slots. The question and the date were the same; asking later means more pages from after the date were online. We saved the pages each answer cited, where the site allowed it, to check their dates; a page we could not save was judged on the detail of the citation.
For scoring, every part of each key was written as a yes-or-no question, 4 to 7 per question. Every answer was scored yes or no on those questions and on five more, and the recheck on the key's questions again. The score weights four parts: Correct (the key's questions) 40, Trustworthy (sources published by the date, the right period, honesty about limits) 30, Usable (answer first, concise) 20 and Second try 10. A critical error scores 0: anything published after the date used in the answer, even when labelled, or a figure attributed to a filing that does not contain it. A missing answer scores 0 too, and a recheck after the time limit counts as none. An assistant's total is the plain average of its six job scores. Equal weights, more weight on correctness, no weight on the recheck and counting every answer alike all leave the order unchanged. Three questions per job is a small base: a job score moves by a third with one answer, and the totals describe these 18 questions, not every question of their kind. Accounts connected to a broker, real money, other plans, repeated runs and investors outside the United States are not covered. How we test personal AI assistants, in full.
This review received no sponsorship of any kind.
Commercial relationships. analyst.ai has no commercial relationship with any product reviewed here: no sponsorship, paid placement, affiliate links, or vendor-provided accounts or credits.
Data. The tables behind this review (every first answer with its yes-or-no counts and score, job scores, and overall and part scores) are available to download from this page.
Corrections. Spot an error? Our corrections log and how to report one: analyst.ai/corrections.
Sources
- Every first answer, one row per assistant and question, with its yes-or-no counts, any critical error and its score (the per-answer download on this page).
- Job and overall scores for each assistant (the score downloads on this page).
- The 18 questions, the answer keys written before testing, and the companies' and funds' own documents published by the date that the keys rest on.
- Each assistant's run records, 26 to 28 October 2026: the question, the full reply, the recheck, the pages it cited and when each answer was given.
- The scoring records, including how disagreements between scorers were settled.