Three Systems, One Prompt: Two Ways to Fail a Source

Framing note: This article compares three single runs on one topic, checked by the same procedure. It is a case study, not a study. Where external research supports the pattern or contradicts it, it appears in the text. It replaces an earlier article that compared only two systems; the predecessor remains available.

Dieser Artikel ist auch auf Deutsch verfügbar: Drei Systeme, ein Prompt

Evidence on file: the unabridged raw reports from Perplexity and ChatGPT, the matching review reports (Perplexity, ChatGPT), the article our own pipeline produced and the predecessor comparing two systems.

One Prompt, Three Systems

Two research tools were given the same question and gave opposite answers. One recommends replacing beef with chicken. The other explicitly advises against it. Both reports read with composure, both cite sources, and from the text alone there is no telling which one is right.

So we looked. Two research tools and an in-house research and verification chain received the identical brief, in the English original:

“Analyze the implications of eating meat for health, animal wellbeing, environmental impact. Come up with the best strategy to achieve the best compromise along those three axes.”

System one: Perplexity Deep Research – a research mode that automatically condenses several search passes into a report with source references – on 29 August 2026, one run on a Pro subscription. The result: roughly 1,400 words with 32 sources, finished in minutes.¹

System two: ChatGPT Deep Research, on 30 August 2026, one run on a Plus subscription. The result: roughly 2,700 words with its own source ledger of fifteen entries, also in minutes.²

System three: our own chain. Five independent research agents – separately running AI instances that know nothing of each other – then an outline, a draft, and a six-stage verification chain. The result: 2,763 words with 35 footnotes, around 20 agent runs, about 1.7 million agent tokens. A token is the unit in which AI text is processed and billed, roughly a word fragment. Duration: one working day.³

All three results then went through the same evidence check. It works with raw extracts: one agent secures the wording of the source character by character, without knowing the claim under review; a second compares claim and extract with no internet access.⁴ Anything that blocks release counts as class A: a false substantive claim, a wrong number or reference base, a false attribution, an undisclosed interest behind a source, anything invented.

The yardstick is not ours. In its own launch post, Perplexity promises “expert-level analysis across a range of complex subject matters” and completion of most research tasks “in under 3 minutes”.⁵ We measure the products against their claim, not against our effort.

And the thesis of this article is: citing is not verifying – and there are two ways to fail at it. You can get a source wrong – reproduce it inaccurately – and you can be mistaken about the source itself: about who compiled it and what it is worth. Anyone publishing under their own name needs a verification step that catches both – including when the reviewer is the one who errs.

The Insight Was There in Minutes

First, the part comparisons like this tend to skip: on substance, all three ended up in almost the same place. Independently, they found the three-axis conflict, the insight that no type of meat wins on health, animal welfare and environment at once, the dividing line between processed and unprocessed meat, and all three landed on “less meat” rather than “different meat”. All three also found the poultry trap: switching from beef to chicken improves the climate footprint but multiplies the number of animals affected per quantity of meat, on the order of two hundredfold.⁶

And then the contradiction from the opening, the hinge of this article: Perplexity makes poultry the primary animal protein source of its recommendation; ChatGPT advises against it.

So the core insight was available in minutes. The reliability in the details was not – and the difference is invisible while reading.

Two Ways to Fail a Source

This is where the check comes in, and its result is not that one report had more errors than the other. It is that the errors are of a different kind.

Perplexity gets the sources wrong. The report reproduces the source inaccurately – not necessarily the underlying facts.

Three examples. It writes that a kilogram of beef requires “roughly 15,400 cubic meters of water” – that would be 15.4 million litres. The cited source says 15,400 cubic metres per tonne, i.e. 15,400 litres per kilo.⁷ The report’s own table carries the correct unit right beside it; text and table contradict each other, and nobody saw it. Second, it claims that the pooled data of six US long-term cohorts did not link poultry to cardiovascular disease. The cited study found the opposite: poultry was statistically linked to disease – an association, not a demonstrated cause: hazard ratio 1.04, a measure of relative risk over time; the study puts the absolute difference at about one percentage point over thirty years.⁸ The effect is small, and the study authors were themselves cautious because of possible preparation effects⁹ – the error lies in the reproduction, not necessarily in the conclusion.

Third, from a UK Biobank study it quotes the most striking sub-finding and omits that the same study found no association with all-cause mortality.¹⁰ In total, five blocking findings remain under our house rules: the three named, plus an estimate presented as fact and one source-quality issue. The count follows our editorial rules and is not an objective measure.

ChatGPT is mistaken about the sources. Here the check found not a single wrong number. Every value examined appears verbatim in its source, including the reference bases. The most serious finding concerns the attribution: the meat yields of 360 kilograms per cow and 1.7 kilograms per chicken, on which the pivotal two-hundred calculation rests, are labelled by the report as “Our World in Data synthesis using FAO/LCA data”. The metadata of the cited page name Faunalytics, an animal-advocacy organisation, as the source of those values.¹¹ The numbers are right; the label upgrades their provenance. “FAO data” reads as UN statistics. Added to that is a second blocking objection – a statement about meeting nutrient needs without meat, backed by a page of the German Nutrition Society that says nothing on the matter – and, among others, a lighter point: three additions to a grazing report whose core is covered while the additions do not appear on the overview page that is linked.

Independent research finds the same profile across this product class. The citation accuracy of Perplexity Deep Research is 90.24 percent in a scientific benchmark.¹² The weaknesses sit deeper: an audit study of this product class measures citation accuracy between 40 and 80 percent alongside pronounced one-sidedness,¹³ another separates link validity from factual accuracy: for the strongest models, link validity stays above 94 percent while factual accuracy reaches only 39 to 77 percent – across all fourteen models tested, factual accuracy ranges from 24 to 77 percent, and Perplexity is not among them.¹⁴ Research now measures precisely the layer at which both reports failed: not whether the link loads, but whether the source supports the sentence.

Both kinds of error share one thing: a working link catches neither. The sources exist, the addresses load, the numbers are mostly there. The break happens on the way from the source to the sentence. Anyone reading this as “one system is the worse one” is reading too fast: the report with the better sourcing discipline is the second – and it is precisely the one carrying the harder finding.

What the ChatGPT Report Does Better Than We Do

The second external report deserves its own section, and not out of politeness.

The check found two blocking findings in it, one of them hard – and not a single wrong number. It marks its own conclusions explicitly as such: “This is a reasoned inference, not a validated scientific ranking across species.” It separates hazard from risk classification and makes clear that the cancer classification of processed meat does not mean it is as dangerous as smoking. It has its own uncertainties section and another one disclosing under which weighting of the three axes the recommendation would flip. It keeps a source ledger with access dates. And on the water footprint it warns explicitly of exactly the trap the other report walked into: “scarcity matters more than liters alone”.²

Our own article does not do two of these things: we do not mark our own inferences throughout, and we keep no separate source ledger – our access dates appear only in the footnotes.

What impresses most is what it did not leave out. On one environmental study it carries over the authors’ caveats about water and biodiversity, even though leaving them out would have made its own statement stronger.¹⁵ That is exactly where the other report failed.

It is not flawless: absolute risks do not appear beside the relative ones here either, interests are not flagged systematically, and several times its references point to overview pages rather than the document that carries the claim.

And still it is this report, of all three, that carries a false label on its most important number. That is the actual proof of the thesis: diligence does not replace verification. The most disciplined of the three texts fails at the layer discipline alone does not cover.

The Look in the Mirror

Anyone who has read this far and concluded that at least one of the two can do it is reading too fast. Our own pipeline produced both kinds of error itself.

We got a source wrong like this: with Perplexity, the study did not support the sentence about poultry – we wrote “sudden cardiac death” into a statement about an EFSA opinion in which the term does not appear.

We were mistaken about a source like this: ChatGPT labelled figures from an advocacy organisation as FAO data – we attributed an agreement rate of 60 percent to the study as a whole, although it holds for only one of its two datasets,¹⁶ and claimed a preference for models of the same family, a point the cited paper explicitly lists as an open question rather than a finding.¹⁷

On the meat article, the verification chain caught sixteen class-A findings, spread across every stage: four in the first quick audit, one in the word-meaning check, seven in the evidence check, three in the strict final audit and one in its second round, a knock-on error of our own correction. On the comparison article itself, eleven more followed.¹⁸ This is not a score to set against theirs. It is the proof that our chain produces the same two kinds of error as the two external systems – it just catches them first.

The difference between the systems is not who makes the errors. It is whose errors reach the reader.

The Most Instructive Case: The Review Makes the Error It Objects To

The most instructive case concerns us.

Our review of the Perplexity text originally listed six findings. Number six: an alleged contradiction, because it says “over 80 billion land animals slaughtered” in one place and “more than 97.6 billion” in another. That sounds like a contradiction, but is not one. 97.6 is greater than 80, both sentences can be true at once, and between the two values sits the estimate an animal-advocacy organisation attributes to the FAO: 92.2 billion land animals kept and slaughtered annually.¹⁹

Not an error in a source and not an error about a source, but an error about our own finding – the third place where citing is not verifying. The reviewer did not catch it. A red team did – a separate agent with the explicit brief to attack our own thesis. The finding was downgraded. Our review had thereby made exactly the error it objected to: a dramatic diagnosis that does not survive its own scrutiny.

It was not the only incident. During the strict final audit, the reviewing agent’s retrieval module invented a date that appears nowhere in the source. What stopped it was not a smarter model but the verbatim raw extract.¹⁸

Verification is not an installation you set up once and then trust. It is a discipline that has to run against the reviewer too, and its most reliable instrument is the simplest one: the verbatim raw text of the source.

What This Comparison Does Not Prove

If even the review errs – what does this comparison prove at all? Less than the headline promises. Here are four limits we would rather write down ourselves.

First: these are three runs. One topic, no repetition, no pre-registered analysis. The case illustrates a pattern that independent research finds across this product class. It does not prove it.

Second: our reviewers share the same blind spots – but the objection has shrunk. In the predecessor article we had to concede that a Claude reviewer may have been checking a Claude report: according to its help centre, Perplexity Deep Research uses “Opus 4.6 Thinking” on the Max subscription and “4.5 Thinking” on Pro,²⁰ model names the trade press attributes to the Claude family.²¹ Our run was on Pro; which model actually served the run is not documented.¹ The ChatGPT report comes from a different model family, not the one our reviewers belong to, and our chain found class-A findings there too. That is a stronger result, not a resolution.

The research picture remains uncomfortable – on one of two datasets examined, wrong answers from model pairs agreed on average 60 percent of the time, against a chance expectation of one third, and it is the more accurate models that correlate more strongly, across provider boundaries as well.¹⁶ When an AI model evaluates answers from various models, it often favours its own – widespread, but not uniform; effects within a model family are explicitly listed by the study as an open question.¹⁷

Third: the comparison was not blind. Our reviewers had a week of topic knowledge from our own article, along with 21 finished raw extracts and a list of known problem numbers. At least one finding was discoverable only because of that.

Fourth: there is a genuine counter-example. In literature search, human-compiled reference lists are clearly outperformed: only 51 percent of human citations were judged moderately relevant or higher by a neutral AI evaluator, against 86 to 88 percent for the strongest AI-based re-rankers. And human authors are 2.5 times more likely than those methods to cite themselves or direct co-authors.²² Our source selection is no reliable yardstick either.

One pass against a different model family is a data point, not an all-clear – our reviewers still all come from the same organisation.

What the Effort Is Worth on Paper – and Why That Sum Was Never Invoiced

Which leaves the question a decision-maker actually asks: does the effort pay off anyway? The cost difference needs to be on the table, correctly labelled.

Perplexity Pro costs 20 US dollars a month, ChatGPT Plus likewise. We could not verify a euro price for either; the amounts circulating in the press are conversions.²³ Both reports were finished in minutes.

Our side cost one working day and about 1.7 million agent tokens. For the day of creation our accounting shows 188 US dollars,²⁴ including other work that day; roughly a third to a half of it was attributable to the article. And that too is not an invoice but a notional equivalent: the amount the same volume of tokens would have cost through the programming interface. None of it was paid. The work ran inside a monthly subscription at a flat rate of about 100 euros.²⁴ The predecessor article called this figure a daily bill; that was wrong, and the correction belongs here.

The gap between the two numbers is itself the finding. For all of August our accounting sums to 1,831.93 US dollars in notional equivalent, or about 1,573 euros after conversion.²⁴ Against a subscription of about 100 euros that is a good fifteen times over. Anyone working this way currently pays a fraction of what the same work costs through the interface.

The distribution inside the chain is more uncomfortable than we would like: verification cost just under three times as much as research, and the single most expensive step was, of all things, the evidence comparison.³ That was down to a fault in the retrieval module – the reviewing agent read a 27,000-line extract in one go instead of searching it. The agent has searched the extract ever since rather than reading it; we have not measured how much that lowers the price. So the price of a verification step is neither fixed nor a technical given: it depends on how the step is implemented and on a vendor decision.

By the measure “insight per euro and minute”, the tools win this comparison clearly. That is stated here in so many words.

For texts appearing under your own name, a different measure applies. A factor-of-1,000 error in an internal research result costs a correction. The same error in a published text costs credibility. By “insight per euro and minute” the tools win. By “errors per signature” they do not.

What follows is not a ranking but a division of labour. Gathering is the tool’s job; it is orders of magnitude faster and cheaper at that, and it delivers the core insight reliably. Evidence checking is the job of a separate verification step – no matter whether the text comes from Perplexity, from ChatGPT, from an in-house pipeline or from a human.

Research Depth Is Now Rationed, and the Research Yardstick Is Shifting

That the price of this division of labour is a vendor decision is not merely an assertion. The movement can be dated.

At launch on 14 February 2025, Perplexity advertised Deep Research as free for everyone, with unlimited runs for Pro subscribers.⁵ In early February 2026 that allowance was cut to about 20 runs per month; the starting figure sits, depending on the source, at 500 to 600 per day, and the cut hit existing customers mid-billing-cycle.²⁵ The reasoning in the help centre reads: “Usage limits have been adjusted to allocate more computing power per session, so you get deeper insights and higher-quality results.”²⁰ Whether quality per run actually rose remains an open question; no independent before-and-after measurement exists.

In parallel, the research yardstick is moving: newer benchmarks increasingly separate “does the link work” from “does the source support the claim”.¹²¹⁴

And the products already differ on this: that the ChatGPT report supplies a source ledger with access dates and an uncertainties section of its own is something the Perplexity report from the same week does not do. That suggests vendors are building in early aids to verifiability, if unevenly – though what is documented here is only a difference between two runs. These aids do not replace verification yet – the false attribution label sits in the same report.

Condensed into three questions before a research result goes out the door:

1. Does every load-bearing source support the sentence that sits on it – checked against the raw text, not against a working link?
2. Is the attribution right too: who compiled the number, what interest do they have, and does the absolute figure stand beside every relative risk?
3. Did the verification also run against the reviewer – with a brief set against your own thesis?

Question one catches the errors in the sources, question two those about the sources. Question three asks both again – of the reviewer. Citing is not verifying: two ways to fail mean two questions for every text that carries your name, and one for whoever checks it.

To close, the transparency offer we would demand of any external report: both raw reports are online verbatim, both review reports as well (Perplexity, ChatGPT), including the finding that did not survive our own scrutiny. The raw extracts of all sources are on file and will be disclosed on request. It is the same verifiability this article demands of others.

Which of the two reports from the opening was right could not be decided from the text. From the raw text it could be decided – in both directions.

The research tool supplies the raw material. Only checking it against the raw extract turns it into something you can put your name to.

Quellen

1. Perplexity Deep Research report of 29 August 2026, one run on a Pro subscription, Deep Research mode, no post-editing. Which model served the run is not documented. The unabridged report is online verbatim on this blog. ↩ ↩²

2. ChatGPT Deep Research report of 30 August 2026, one run on a Plus subscription, no post-editing. The unabridged report is online verbatim on this blog, as is the full review report with all findings. The word count is our own count of the raw report including its source ledger; the ledger contains fifteen entries. ↩ ↩²

3. “Not Which Meat, but How Much” (A16), this blog – the article our pipeline produced from the same prompt. Effort figures from internal run and verification logs; disclosed on request. ↩ ↩²

4. Methodology reference: two-stage evidence check. One agent secures the wording of the source without knowing the claim under review; a second compares claim and extract with no internet access. Prompt templates in the open repository: https://github.com/flodido/agentic-research-vault ↩

5. Perplexity: “Introducing Perplexity Deep Research”, 14 February 2025. Original URL https://www.perplexity.ai/hub/blog/introducing-perplexity-deep-research (blocks automated access); checked verbatim against the Wayback snapshot from launch day: https://web.archive.org/web/20250214203504/https://www.perplexity.ai/hub/blog/introducing-perplexity-deep-research (accessed 30 August 2026). Vendor self-description, accordingly interest-driven. ↩ ↩²

6. Bryant Research: “The Small Body Problem”, https://bryantresearch.co.uk/insight-items/small-body-problem/ (raw extract of 30 August 2026): “about 200 chickens are required to produce the same amount of meat gained from slaughtering one cow”; for the same quantity of meat the source itself names the alternative values 150:1 and 192:1, whereas its 223:1 refers to the annual total slaughtered rather than equal meat quantity. Ritchie, H.: “What are the trade-offs between animal welfare and the environmental impact of meat?”, Our World in Data, 10 June 2024, https://ourworldindata.org/what-are-the-trade-offs-between-animal-welfare-and-the-environmental-impact-of-meat (raw extract of 30 August 2026): 360 kg per cow, 1.7 kg per chicken, “200 times as many chickens as cows”. Calculated from the yields, this gives 212. Carried in the text as an order of magnitude. ↩

7. Water Footprint Network: “What can consumers do?”, https://www.waterfootprint.org/time-for-action/what-can-consumers-do/ (raw extract of 30 August 2026). There: 15,400 m³/t for beef. ↩

8. Zhong, V. W. et al.: “Associations of Processed Meat, Unprocessed Red Meat, Poultry, or Fish Intake With Incident Cardiovascular Disease and All-Cause Mortality”, JAMA Internal Medicine, 3 February 2020, https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2759737 (raw extract of 30 August 2026). Poultry: for each additional 2 servings per week, HR 1.04 (95% CI 1.01–1.06) for incident cardiovascular disease, 30-year absolute risk difference 1.03 percentage points; no significant association with all-cause mortality. ↩

9. Northwestern Feinberg School of Medicine (press office of the study institution): “Meat Consumption Raises Risk of Heart Disease and Death”, 3 February 2020, https://news.feinberg.northwestern.edu/2020/02/03/meat-consumption-raises-risk-of-heart-disease-and-death/ (accessed 30 August 2026): “the evidence so far is not sufficient to make a clear recommendation about poultry intake.” ↩

10. “Red meat consumption and all-cause and cardiovascular mortality: results from the UK Biobank study”, European Journal of Nutrition, DOI 10.1007/s00394-022-02807-0; UK Biobank publication overview: https://www.ukbiobank.ac.uk/publications/red-meat-consumption-and-all-cause-and-cardiovascular-mortality-results-from-the-uk-biobank-study/ (raw extract of 30 August 2026). The study found elevated cardiovascular sub-risks, “but not all-cause mortality”. ↩

11. Chart metadata of the Our World in Data page named in footnote 6, dataset “kilograms-meat-per-animal”: source “Faunalytics (2023) – Impact of Replacing Animal Products”, an advocacy organisation. The FAO is not named for these values. According to our review report the emissions data in the same chart come from Poore and Nemecek (2018); that addition is not contained in the metadata extracts available to us. ↩

12. DeepResearch Bench: “A Comprehensive Benchmark for Deep Research Agents” (arXiv:2506.11763), https://deepresearch-bench.github.io/ (accessed 30 August 2026). FACT framework, 100 PhD-level tasks across 22 fields: 90.24 percent citation accuracy for Perplexity Deep Research. ↩ ↩²

13. Venkit, Laban et al. (Salesforce AI Research): “DeepTRACE: Auditing Deep Research AI Systems for Tracking Reliability Across Citations and Evidence”, https://arxiv.org/abs/2509.04499 (accessed 30 August 2026). Note the vendor proximity: Salesforce runs an agent business of its own. ↩

14. “Cited but Not Verified: Parsing and Evaluating Source Attribution in LLM Deep Research Agents” (arXiv:2605.06635), https://arxiv.org/html/2605.06635 (accessed 30 August 2026). Verbatim: “even the strongest frontier models maintain link validity above 94% and relevance above 80%, yet achieve only 39–77% factual accuracy”; across all fourteen models tested, factual accuracy ranges from 24 percent (OSS-120B) to 77 percent (Claude Opus 4.5). Perplexity is not among them. Complementary on the product class: “DEER” (arXiv:2512.17776) and DeepResearch Bench II (arXiv:2601.08536), both accessed 30 August 2026 – even the strongest agents remain clearly behind expert grading rubrics. ↩ ↩²

15. Scarborough, P. et al.: “Vegans, vegetarians, fish-eaters and meat-eaters in the UK show discrepant environmental impacts”, Nature Food, 20 July 2023, https://www.nature.com/articles/s43016-023-00795-w (raw extract of 30 August 2026). The report carries over both the core findings and the authors’ caveat about water and biodiversity. ↩

16. Kim, Garg, Peng, Garg: “Correlated Errors in Large Language Models”, ICML 2025, https://arxiv.org/abs/2506.07962 (accessed 30 August 2026). Over 350 models across two leaderboards and a résumé-screening task. The 60 percent value holds for the Helm dataset with 71 models; on the second leaderboard the mean is 42.3 percent. The paper itself states the chance expectation of one third for the Helm dataset: “choosing between incorrect answers uniformly at random would lead to an agreement rate of 1/3”. More accurate models correlate more strongly “even with distinct architectures and providers”. ↩ ↩²

17. “Quantifying and Mitigating Self-Preference Bias of LLM Judges”, https://arxiv.org/abs/2604.22891 (accessed 30 August 2026): widespread, but not uniformly self-favouring – of 20 models, 8 biased positively, 9 negatively, 3 neutral. Family effects across model boundaries are explicitly named by the paper as an open, unresolved limitation. ↩ ↩²

18. Verification logs of the two preceding articles; disclosed on request. They contain the stage-by-stage breakdown of the findings as well as the documented hallucination of the retrieval module in the strict final audit. ↩ ↩²

19. Humane World for Animals: “More animals than ever before – 92.2 billion – are used and killed each year for food”, https://www.humaneworld.org/en/blog/92-billion-animals-used-killed-meat-each-year (accessed 30 August 2026). NGO source with an advocacy interest; used here solely as evidence of an estimate sitting between the two values, which the organisation gives without a reference year. ↩

20. Perplexity Help Center: “What’s New in Advanced Deep Research”. Original URL blocks automated access; checked verbatim against the Wayback snapshot of 11 February 2026: https://web.archive.org/web/20260211012137/https://www.perplexity.ai/help-center/en/articles/13600190-what-s-new-in-advanced-deep-research (accessed 30 August 2026). The page carries no absolute modification date. ↩ ↩²

21. Dataconomy: “Perplexity Upgrades Deep Research Tool With Claude Opus 4.5 Integration”, 5 February 2026, https://dataconomy.com/2026/02/05/perplexity-upgrades-deep-research-tool-with-claude-opus-4-5-integration/ (accessed 30 August 2026). ↩

22. Sahu, Charlin, Pal: “Rethinking Literature Search Evaluation: Deep Research Helps, and Human Citation Lists Are Not a Ground Truth”, https://arxiv.org/pdf/2605.29234 (accessed 30 August 2026). Verbatim: “only 51% of human citations are judged moderately relevant or higher, against 86–88% for the strongest AI-based re-rankers” and, in the body of the study, “At d=0 (direct co-authorship) humans are 2.5× more likely than the strongest re-rankers to cite themselves or a co-author (5.13% vs. 1.9–2.1%)”. For clarity: the recall improvement from below 20 to above 80 percent also reported in the study measures something else, namely its own pipeline against plain API search, and is deliberately not used here. ↩

23. OpenAI: “What is ChatGPT Plus?”, Help Center, https://help.openai.com/en/articles/6950777-what-is-chatgpt-plus, checked against the Wayback snapshot of 22 August 2026 (accessed 30 August 2026): “ChatGPT Plus is a subscription plan that provides enhanced access to the ChatGPT web app for $20/month.” Neither there nor on OpenAI’s pricing page (last usable archive state of 4 September 2025) is a euro price or a VAT note given. The price of Perplexity Pro is documented by the limit tracker in footnote 25, which lists “Pro ($20/mo)”; we did not extract a Perplexity pricing page, and no euro price is known to us there. ↩

24. Our own evaluation of local usage logs with the accounting tool ccusage, as of 30 August 2026, 20:51; raw output archived. The tool computes what the tokens consumed would have cost through the programming interface – it is expressly not an invoice. August 2026: 1,831.93 US dollars in notional equivalent. Conversion at the ECB reference rate of 28 August 2026 (0.85889 EUR/USD) gives about 1,573 euros. The subscription price of about 100 euros a month is the account holder’s own statement. The ratio is meaningful only after conversion to a single currency. ↩ ↩² ↩³

25. MakeUseOf: “If you bought an annual Perplexity subscription, you were lied to”, February 2026, https://www.makeuseof.com/bought-annual-perplexity-subscription-lied/ – the author is himself an affected annual subscriber (“including myself”); the reduction figures are independently confirmed by the limit tracker verified against the Wayback Machine, https://ailimit.watch/tools/perplexity/ (as of 6 June 2026, accessed 30 August 2026), which flags the cut as unannounced. ↩