Log 001 asked four AI engines what Reforge is, using the company name. This asks the same question using only the URL, and compares the answers.
Bottom line: giving an AI engine your URL instead of your company name doesn’t reliably change what it says about you. What did change the answer was a six-year-old copy of the company’s website, still live on a hostname nobody was watching
Table of Contents
Run 1 — 13 August 2026
The Query
Asked verbatim, in a fresh session on each engine: What is reforge.com, and what are they best known for?
One word different from Log 001. The company name is gone; the URL replaces it. Nothing else changed.
Test Conditions
- Log reference ID: AIVL-002
- Topic category: AI Retrieval & Source Selection
- Target subject: reforge.com (URL only, no company name supplied)
- Engines tested: Claude (Anthropic), ChatGPT (OpenAI), Gemini (Google), Perplexity
- Models: Claude: Sonnet 5. Gemini: Flash. ChatGPT and Perplexity: auto.
- Test environment: Clean test account. Personalization surfaces verified individually. Fresh session per run.
- Runs per engine: 1
- Date of runs: 13 August 2026
- Baseline: Log 001, same question by name
- Status: Single-run. Three further runs planned; findings here are directional.
What Changed
The test case is a fact with a date on it. Miro acquired Reforge in March 2026, and Reforge’s own site carries the announcement. Any engine describing the company as independent is working from a picture more than four months out of date.
Log 001 asked about Reforge by name, across four rounds. Two engines reported the acquisition consistently; two never did. Here is the same measure under the URL query.
| Engine | Acquisition reported, by name (4 runs) | Acquisition reported, by URL (1 run) |
|---|---|---|
| Claude | 4 of 4 | Yes |
| ChatGPT | 1 of 3 | No |
| Gemini | 0 of 4 | No |
| Perplexity | 0 of 4 | Yes |
Perplexity moved. Gemini didn’t. The other two held their existing positions.
One thing didn’t move anywhere. The URL changed which sources were retrieved, but it did not change how any engine described the company. Category, audience and subject matter were identical to Log 001 across all four. What a URL appears to affect is where an engine looks, not what it concludes about the basics.
Sources Cited
| Engine | Citations | Unique pages | Reached corporate news? | Sources |
|---|---|---|---|---|
| Claude | 9 | 5 | No | Tracxn · reforge.com · Mind the Beet · CB Insights (×3) · Chrome Web Store listing (link broken) |
| ChatGPT | 5 | 4 | No | reforge.com · product specializations blog post · reforge.com/learning · a Reddit thread |
| Gemini | 7 | 5 | No | reforge.com (×2) · /teams (×2) · Growth Series detail page · product management category (×2) |
| Perplexity | 8 cited, from 15 retrieved | 8 | Yes | Miro newsroom · reforge.com · Preqin · businessmodelcanvastemplate · a LinkedIn programme review · a Reddit thread · PRNewswire Series B release · Chrome Web Store listing |
The fourth column is the whole result. One engine reached a page announcing the acquisition. It is the one engine whose answer changed.
Perplexity’s full 15 retrieved results — 13 August
Chrome Web Store (Reforge Build) · Miro newsroom · Preqin · PitchBook · Tracxn · reforge.com · Reddit · Business Model Canvas Template · LinkedIn company page · Mind the Beet · LinkedIn programme review · Toption · Vizologi · PRNewswire · Vendr
Compare against Log 001’s fifteen, which included ZoomInfo, Owler, Built In, SignalHire, Levels.fyi, Craft, VentureCapitalTracker, the a16z announcement, Startup Anatomy, and a Tracxn profile for an unrelated company.
Claude: reported it, without citing it
Claude reported the acquisition, as it has in every run. Its citations: Tracxn, the homepage, a personal upskilling blog, CB Insights three times, and the Chrome Web Store listing for Reforge Build.
No Miro newsroom. No news page. No about page. Claude has now reported this acquisition in five runs without once citing a page that announces it. CB Insights tracks ownership changes and is the plausible route, but the interface doesn’t say so, so Claude remains the one engine whose mechanism can’t be inspected.
ChatGPT: no corporate page, no acquisition
ChatGPT cited the homepage, a blog post on product specializations, /learning, and the same Reddit thread that has appeared in every ChatGPT run of this series. No corporate page, no acquisition: consistent with its clean-run behaviour in Log 001, where the acquisition appeared only in the run that reached the founder profile.
Gemini: no movement at all
Seven citations, five unique pages: the homepage twice, /teams twice, the Growth Series course detail page, and a product-management course category page twice.
Course pages and the teams page. The same corner of the site it retrieved in every run of Log 001.
Handed the URL explicitly, Gemini went back to the catalogue and missed the acquisition again. That is now five consecutive runs across two weeks and two different phrasings, describing an acquired company as independent.
So the URL is not a lever that works on every engine. On this one it appears to change nothing.
Perplexity: a different corpus
Perplexity retrieved fifteen results for both queries. Almost none of them overlap.
Gone under the URL query: ZoomInfo, Owler, Built In, SignalHire, Levels.fyi, Craft, VentureCapitalTracker, the a16z funding announcement, Startup Anatomy — and a Tracxn profile for an unrelated company that also happens to be called Reforge.
New under the URL query: Miro’s own newsroom announcement of the acquisition, Preqin, the Chrome Web Store listing for Reforge Build, two independent programme reviews, a pricing aggregator, and a business-model breakdown.
Asked by name, Perplexity assembled an employer profile: headcount, salaries, funding rounds, who works there. Asked by URL, it assembled a product picture: what the thing is, what it costs, whether it’s any good, and who owns it now.
The acquisition announcement was never in the pool for the name query. Under the URL query, it was retrieved and cited first. The answer changed because the corpus changed.
The duplicate company also disappeared. In Log 001, Perplexity pulled a profile for a different business sharing the name. A URL is unambiguous in a way a company name isn’t. For anyone whose brand name is also a common word or a competitor’s, that’s the most directly useful line in this log.
Why the miss happens: the pages don't say it
After running these tests I went and read reforge.com the way an engine would.
Miro appears nowhere in the homepage’s substance. Not in the title tag, which reads Reforge | Product Management Courses for Individuals & Teams. Not in the meta description. Not in the headline, which is about AI rewriting product. Not in any body copy, pricing table, course listing or testimonial.
It appears in three places, all in the footer: a Careers link pointing at miro.com, the Terms and Privacy links pointing at miro.com, and one link labelled Reforge x Miro going to a blog post announcing the deal.
Gemini and ChatGPT aren’t reading badly. The pages they read genuinely don’t carry the fact. Course pages, category pages, the teams page, the homepage: none of them mention the acquisition, because Reforge kept its name (after acquisition) and its product line, so none of them needed rewriting.
The likely reason the whole thing is invisible: nothing changed on the surface. A company that rebrands after an acquisition updates every page automatically, and every engine picks it up. A company that keeps its name updates one blog post, and the news lives or dies on whether an engine happens to find it.
This is the most portable lesson in the log: If your most significant recent fact exists on one page that nothing prominent links to, engines reading your main pages will describe the company you used to be, and they will keep doing it, consistently, for months.
One clear error
Claude stated that Reforge Build launched since the Miro acquisition.
It didn’t. Reforge introduced Build on 12 November 2025, four months before the deal. Miro’s announcement lists Build alongside Insights and Research as products Reforge already had. Perplexity, citing the same Chrome Web Store page, described Build as pre-acquisition and was right.
Reforge’s own announcement post is worth a note. It is still fully indexed by search engines, with its title, date and opening paragraphs intact, but the page itself now returns a 404. The surviving public record of the launch is third-party: a post by Ravi Mehta, one of the product’s creators, published the same day.
The likely mechanism is worth naming. The Chrome Web Store listing says what Build is. It doesn’t say when it shipped. Claude appears to have supplied the timing itself, ordering it against the acquisition it already knew about, and got the sequence backwards.
That’s a specific failure mode: the source establishes existence, the engine infers chronology, and the inference sounds plausible enough that nobody checks.
Two display faults
Claude’s citation list contained two malformed URLs: wwww.reforge.com, with an extra character, and a Chrome Web Store link resolving to an error page.
Neither is an invented source. Both are broken renderings of real ones. But the Build claim- the one that turned out to be wrong- rests on the link that doesn’t open. A reader trying to check it can’t.
Fact-check
- The acquisition was announced on 24 March 2026. Correct in Claude and Perplexity.
- Reforge Build launched on 12 November 2025, before the acquisition. Claude wrong; Perplexity right.
- Balfour’s own statement puts the alumni community above 100,000. Perplexity’s “50,000–100,000+” spans it.
- No engine mentioned Reforge Launch. Class Central reports five AI products shipped in nine months: Build, Insights, Research, Launch and Artifacts.
What this means if you are the brand being asked about
How the question is phrased decides which version of you gets described. Same company, same day, same engine. Asked one way, Perplexity returned a stale funding profile. Asked another, it returned current corporate news. Your prospects use both phrasings and you control neither.
A URL is a disambiguator. If your company name is shared, generic, or close to a competitor’s, name-based queries risk pulling somebody else’s data into your answer. That happened in Log 001 and stopped happening here.
But you cannot rely on it. Gemini ignored the URL entirely and stayed in the product catalogue. Whatever a URL does, it does per engine, and one of the four here did not respond to it at all.
And an engine that can’t reach your news will invent the timeline. Claude knew about the acquisition, knew Build existed, and constructed a relationship between them that isn’t true. Facts that exist in isolation get connected by inference, and the inference is not marked as one.
Limitations - Run 1
- Perplexity’s search depth may have differed. A notification appeared during this run indicating a Pro model or Pro search was active. Perplexity’s free tier includes a limited allowance of these, and they run a deeper retrieval process than the standard mode. If the Log 001 name-query runs were standard searches and this one was not, then the comparison is between two different retrieval processes as well as two phrasings, and the corpus shift cannot be attributed to the URL alone. The interface does not record which mode ran after the fact.
- One run per engine. Log 001 showed how much a single run can mislead: two findings there did not survive repetition. Three further runs are planned, and the figures here may move.
- Six days separate this from the Log 001 baseline. Perplexity’s shift could be the URL, or an index that changed in the interim. A same-day control on the name query is needed to separate them, and this round doesn’t have one.
- One subject, one industry. A company with a thin site or no database listings would likely behave differently.
- All four engines retrieved. This is a retrieval test throughout, not a measure of unaided recall.
- Publishing is a feedback loop. This log is now itself a page about Reforge, and engines may retrieve it in future runs. The measurement becomes part of what it measures.
Runs 2 to 4 — 20 August 2026
Three runs per engine, one week after the first. The single run in the first round suggested the URL had changed one engine’s retrieval almost entirely. Three runs say something different, and the correction is the point of this round.
Test Conditions
- Log reference ID: AIVL-002.1
- Query: Unchanged from Run 1
- Runs per engine: 3 (12 total)
- Date of runs: 20 August 2026
- Models: Claude: Sonnet 5. Gemini: Flash. ChatGPT and Perplexity: auto.
- Test environment: Clean test account. Personalization surfaces verified individually before each session. Fresh session per run.
The acquisition, across both conditions
| Engine | Asked by name | Asked by URL |
|---|---|---|
| Claude | 4 of 4 | 3 of 3 |
| ChatGPT | 1 of 3 | 0 of 3 |
| Gemini | 0 of 4 | 0 of 3 |
| Perplexity | 0 of 4 | 1 of 3 |
The URL made no reliable difference. One engine got slightly worse, one got slightly better, and the two at the extremes didn’t move at all. Whatever effect the phrasing has is smaller than the variation between runs of the same phrasing.
That is a correction to what the first round appeared to show, and it is the reason the first round wasn’t published as a finding.
Sources cited — 20 August
| Engine | Run | Sources displayed | Acquisition reported |
|---|---|---|---|
| Claude | 1 | Tracxn (×3) · Mind the Beet · PitchBook · CB Insights (×2) | Yes |
| Claude | 2 | Tracxn · PitchBook · RocketReach · Chrome Web Store listing (resolves to an error page) · CB Insights | Yes |
| Claude | 3 | wwww.reforge.com (×2) — the 2020 site · Mind the Beet (×2) · Chrome Web Store listing (error page) · CB Insights (×2) | Yes |
| ChatGPT | 1 | reforge.com · Growth Foundations · courses index · learning · AI Growth course | No |
| ChatGPT | 2 | Growth Foundations · learning · courses index · AI Growth course | No |
| ChatGPT | 3 | reforge.com · $21m funding post · courses index · product-led growth post | No |
| Gemini | 1 | techscaler.co.uk · Tracxn — no Reforge page at all | No |
| Gemini | 2 | None displayed | No |
| Gemini | 3 | None displayed | No |
| Perplexity | 1 | Miro newsroom (×7) · PitchBook · Preqin · businessmodelcanvastemplate · Reddit · LinkedIn · reforge.com | Yes |
| Perplexity | 2 | reforge.com · Preqin · LinkedIn · coffeeandstudy · coursefacts · toption | No |
| Perplexity | 3 | reforge.com · courses index · teams · Mastering Product Management · uxcel | No |
Every source each engine displayed, run by run. Pooling these by engine would hide the pattern. Perplexity’s three runs are not variations on one source set, they are two different source sets.
Claude: unchanged under both phrasings
Three of three under the URL query, four of four under the name query. Eight consecutive runs across three weeks and two phrasings, reporting the acquisition every time. Founding year, founder and headquarters were identical across all eight. The URL changed one thing: Andrew Chen dropped out entirely, credited in none of these three runs after two of four under the name query.
Claude displayed citations in all three runs, and CB Insights appeared in every one. It’s a database that tracks ownership changes, and it is the most plausible explanation for the one engine that has never missed the acquisition despite never citing a page that announces it.
The rest of its sourcing was third-party: Tracxn, PitchBook, RocketReach, a personal blog, and a Chrome Web Store listing that resolves to an error page in both rounds where it appeared.
Across three runs asking specifically about reforge.com, Claude reached a page on that domain exactly once, and it was the wrong one. That finding has its own section below.
One error: run 2 named the customer-feedback product “Insight Analytics.” It is Reforge Insights.
ChatGPT went the other way
Under the name query, ChatGPT reached Brian Balfour’s profile page in one run of three — and that was the run that reported the acquisition.
Under the URL query, it reached that page in none. Every citation across all three runs was a course, catalogue or learning page: the homepage, /courses, /courses/growth-foundations, /learning, /courses/ai-growth, /courses/ai-product-leadership, and two blog posts.
Naming the domain pushed it further into the product catalogue, not toward the corporate pages. Which is where Gemini has been the whole time.
Two side effects worth recording:
Andrew Chen returned. Credited as Growth Series co-creator in all three URL runs, after none of three under the name query. He appears when the growth course and funding-announcement pages get retrieved, which is where that history lives.
Third-party commentary disappeared. The same Reddit thread on whether Reforge is worth $2,000 appeared in all three name-query runs and none of the URL runs. Asked by domain, ChatGPT returned the company’s self-description with no outside voice in it at all.
Gemini: same answer, different sources
Eight runs. Two phrasings. Three weeks. The acquisition has never appeared. It is the most stable result in this series, and it is a wrong answer.
But its sourcing did move. Across every earlier run, Gemini read Reforge’s own catalogue: course pages, category pages, the teams page. Asked by URL, run 1 cited two sources: a UK training aggregator and Tracxn. No Reforge page at all. Runs 2 and 3 displayed no citations, so nothing can be said about them.
Handed the domain, the one Gemini run whose sourcing is visible left the domain entirely.
Perplexity: what the first round got wrong
The single URL run on 13 August had Perplexity retrieving Miro’s newsroom announcement and reporting the acquisition, against a baseline of nought from four name-query runs. The obvious reading was that naming the domain pulled it out of the investor databases and into corporate news.
Three runs show it lands in one of two source sets, seemingly at random:
- One run led with Miro’s newsroom page, cited it seven times, and named all four AI products — Insights, Research, Build and Launch. The only run in twenty-four to mention Launch at all.
- Two runs read the review-site cluster instead: coffeeandstudy, coursefacts, uxcel, toption, plus LinkedIn and Preqin. Neither mentioned the acquisition.
Same query. Same day. Same account. The 13 August run wasn’t the URL working. It was the same one-in-three draw, and a single run couldn’t tell the difference.
The 2020 website that is still live
Claude kept describing Reforge in language I couldn’t find anywhere on their site. “Career Accelerator Programs.” An “intensive 6-week format.” “Results over certificates.” “Alumni have gone on to build many of the world’s most successful companies.” None of that is current positioning.
Run 3 explained it. Two of its citations pointed at reforge.com with an extra character in the hostname.
That address resolves. It serves a complete copy of Reforge’s website from 2020: old logo, old design system, Career Accelerator positioning, Fall 2020 programme dates, and an application deadline of 7 August 2020. Every phrase Claude used that I couldn’t place is on that page.
Nothing appears to link to it. DNS answers for the typo, the certificate doesn’t cover it, and the site behind it was never taken down. So Claude wasn’t inventing anything. It read a real page on a hostname the company owns and reported what was there. The page is six years old.
It gave no indication the source was dated. No caveat, no hedge. The 2020 positioning arrived in the same register as everything else.
Two things follow. The first is narrow: across three runs asking about reforge.com, the only time Claude reached the company’s own domain, it landed on an abandoned copy rather than the live site.
The second is general, and it is the most useful thing in this log. What an engine says about you is bounded by what of yours is reachable, including the things you forgot were reachable. A migration leaves hosts running. A platform change orphans old posts. Nobody audits what still answers, because nothing links to it and no analytics record it. An engine has no such blind spot. It finds whatever responds and treats it as the company speaking.
Worth checking on your own domain: typo’d subdomains, staging, dev, old, beta, www2, and any platform subdomain left over from a migration. A certificate transparency search will list every hostname that has ever had a certificate issued against your domain.
Two errors that recurred
“Reforge Insight Analytics.” The product is Reforge Insights. Claude produced the wrong name on 13 August; Perplexity produced the identical wrong name on 20 August. Two engines, a week apart, converging on the same malformed product name is more interesting than either instance alone.
“Playbooks” as a brand term. Gemini has now presented playbooks in quotation marks, as though it were Reforge’s own vocabulary, in three of eight runs, while naming Artifacts, the actual branded term, correctly in the same responses. Log 001 recorded this as a one-off after it failed to recur. Three of eight is a pattern.
What this adds for a brand
Giving an engine your URL is not a lever. It was a reasonable hypothesis: name the domain, get the site read, get current information. Across twelve runs it produced no consistent improvement, and for one engine it produced the opposite.
What decides the answer is which of your pages get retrieved, and that varies run to run. The same engine asked the same question on the same day reached corporate news once and review sites twice. A prospect asking about your company is drawing from that distribution, not from a fixed answer.
Audit what of yours is still reachable. The most consequential thing in this round wasn’t the query phrasing. It was a six-year-old copy of a company’s site, live on a hostname nobody was watching, being read and quoted as current. Nothing linked to it and no analytics would have shown it.
Which means a single check tells you very little. If you have ever asked an engine about your own business and been satisfied with the answer, that was one draw. Three runs would probably have shown you something else.
Limitations - Run 2
- Three runs separates stable from unstable. It does not quantify either. 0 of 3 and 1 of 3 are not meaningfully different from each other.
- Perplexity’s search mode was not recorded per run. A Pro-search notification appeared during the first round. Whether it fired in these runs is unknown, and it may be the actual variable behind the two source sets.
- Gemini displayed citations in one run of three. Runs 2 and 3 showed none. I did not ask the model to list its sources afterwards, since a retrospective account is a self-report rather than a record.
- Two conditions, one subject. A company whose corporate news sits deeper in the site than its product catalogue. A differently structured site would likely behave differently.
What's next for this log
The URL query has now been run four times across two dates. The answer is that it makes no reliable difference to what engines report.
One gap remains. The name-query figures come from 6 and 12 August, so the comparison is spread across three weeks rather than run side by side on a single day. That control is still worth running, though it matters less than it did: the finding rests on Perplexity returning two different source sets for the identical query on the same day, and no time gap touches that.
The subdomain finding is the more consequential result, and it points somewhere else entirely. What a company leaves reachable is a bigger variable than how a prospect phrases the question.
Revision history
13 August 2026 — Published. One run per engine, URL only.
20 August 2026 — Added three runs per engine. The first round’s central finding did not survive: Perplexity’s corpus shift was run variance, not an effect of the URL. Across twelve runs the URL query produced no reliable change in whether engines report the acquisition, and for two engines it reduced retrieval of the company’s own current pages. This round also surfaced a live 2020 copy of Reforge’s website on a typo’d hostname, cited by Claude and described as current. Original 13 August results retained above.
Part of the AI Visibility Logs, an open record of how AI systems find, interpret, and describe brands. How AI Visibility tests are run.
