The price is wrong: how often AI answers quote the current price of software in India

01 / 08Summary

A buyer who asks an AI engine what a piece of software costs gets a confident list of plans and rupee figures. We checked those figures. On 5 October 2026 we asked ChatGPT (search on), Perplexity, Gemini and Google AI Mode what 30 software products widely bought by Indian small businesses cost in India, and graded every rupee price in the 360 answers against the vendor’s own pricing page, captured from India the same day.

The headline is not that the engines invent numbers from nothing. Most single prices were right. The trouble is that almost every answer quotes several, and one stale or misread figure was enough to mislead.

02 / 08Why this matters

Price is the question a buyer asks once they have a shortlist, and it is a question with a checkable answer. A vendor publishes a price list; an answer either repeats it or it does not. That makes price a clean test of something harder to measure elsewhere: whether an engine is reading the source it should, and reading it now.

For Indian buyers the test is harder than it looks. Many vendors show different prices by country, offer monthly, quarterly and annual billing, run introductory discounts, and print a struck-through list price beside the price you pay. Prices also move: Microsoft, Swipe and Petpooja all showed different rupee prices on their own pages at some point since January 2024. An answer can be right last year and wrong today.

Most single prices were right. Most answers quote several, and one bad figure is enough.

Study 004, Finding 1

03 / 08Method

We started with the facts. On the day of the study, just before and during collection, from an Indian connection, we captured each vendor’s official pricing page, switching every billing toggle the page offers (monthly, quarterly, annual, device, term), and recorded every plan, every rupee price, its billing basis and its tax note.¹ We dropped products whose official page does not show a verifiable rupee plan list, and replaced them with products that do.² The result is a price book of 274 rupee prices across 124 plans.

Each product was asked about in two phrasings, “How much does X cost per month in India?” and “X pricing plans in India”, on four engines, with the request sent from India.³ To test stability, half the products were asked again the same day.

Answers

360240 in the main run, 120 in the repeat run

Products

302 phrasings × 4 engines each

Prices graded

1,111paid plan prices in the main run, about 4.8 per answer

Engines

ChatGPT (search on), Perplexity, Gemini, Google AI Mode

Market

India; requests sent from India; prompts named India as a buyer would

Fieldwork

5 October 2026, 11:33–12:04 UTC; pricing pages re-checked after the last answer, unchanged

How a price was graded

We listed every price each answer gave for a paid plan of the product, with the plan name, the amount and the period it covers. Add-ons, extra users, per-message fees, onboarding fees, other products, and totals the answer worked out itself (team totals, prices with GST added) were left out.⁴ Each listed price was then graded by code against the price book:

• Current: within 5% of a price the vendor’s page shows today for that plan, on any billing basis, after putting monthly, quarterly and annual figures on the same footing. A struck-through list price printed on the page counts as current; the engine repeated what the page shows.
• Outdated: not current, but within 5% of a price the vendor itself published for that plan earlier, found in archived copies of its pricing page or still printed on another page of its site.⁵
• Wrong: neither.

An answer is correct only if every rupee plan price in it is current. It is outdated if at least one price is outdated and none is wrong, and wrong if any price is wrong. Answers that gave prices only in US dollars, gave no figure, or whose captured text visibly lost a price table are counted separately.⁶ Prices for tiers the vendor does not publish (“contact sales”) could not be checked and were set aside.⁷

A citation is a link the engine presented as a source; each URL counts once per answer. A vendor’s pricing page is a page on the vendor’s own site whose address is a pricing, plans or buy page, not a blog post or help article.⁸

04 / 08Findings

1. Fewer than six in ten answers got every price right

Of the 240 answers in the main run, 232 quoted at least one rupee price. 133 of them (57%) had every price match the vendor’s page. 79 answers contained at least one price that was neither current nor a price the vendor had published before. Only 20 were off solely because they repeated an older official price. Of the 225 prices that were not current, 67 (30%) matched a price the vendor itself had published before; the rest we could not trace to the vendor.

Counted price by price, the picture is better: 886 of 1,111 prices (80%) were current, 67 (6%) were older official prices and 158 (14%) were wrong. The gap between the two views is the length of the answers. A typical answer quoted about 4.8 prices; answers with one or two prices were fully right in 33 of 45 cases, those with six to nine in 41 of 78.

2. The cheapest plan was misquoted in one answer in four

Buyers usually start from the entry price. We took the lowest-priced paid plan each answer named and graded it on its own: 176 of 232 (76%) were current, 24 were an older official price and 32 were wrong.

3. Engines differ, by less than the products do

ChatGPT had the highest share of fully correct answers, 66% of 59; Perplexity the lowest, 50% of 56. On the entry price alone the four engines ranged from 71% to 80%. Across products the spread was far wider: from all eight answers fully right to none.

4. Misses were mostly modest, and sometimes not

For 136 misquoted prices we could set the figure against a current price of the same plan, on a monthly basis, choosing the closest one.⁹ The median gap was 19%. 78 were too high and 58 too low. For per-user plans the median gap was ₹295 per user per month (n = 62), which for a team of ten is a different budget line.

5. Citing the pricing page helped, but did not settle it

177 of 240 answers (74%) cited the vendor’s own pricing page. Those answers were fully right 62% of the time. Answers that cited no vendor page at all were fully right 50% of the time. So 38% of answers that cited the pricing page still had at least one price that was not current. Gemini cited the pricing page least often, in 32% of its answers, against 92% for ChatGPT.

6. “No rupee prices here”

22 of the 360 answers told the buyer that the vendor does not publish a rupee price, or prices and bills only in US dollars. The claims covered 8 products. From an Indian connection, every one of those vendors’ pricing pages showed rupee prices on the day of the study.

Several of these vendors show prices by the visitor’s country. Archived copies of their pages, captured from outside India, show dollars where an Indian visitor sees rupees. 9 answers in the main run gave at least one plan price in US dollars; in 4 of them dollars were the only price given.

7. Some products were answered well every time

Accuracy depended more on the product than on the engine. All eight answers were fully right for TallyPrime, Kylas and Zoho Payroll; for Swipe, none of the eight was. The products answered worst include several that renamed plans or changed prices in the past two years, and several whose pricing page shows different prices by country. 58 of the 1,111 prices were given under a plan name that is not on the vendor’s page today, across 9 products. The clearest case of stale prices was WATI: 6 of its 8 answers repeated prices from its own earlier pricing page and nothing worse. Answers that were not fully right drew a slightly larger share of their citations from review and listing sites, 41% against 37%.

8. Ask again, get a different verdict one time in five

We asked the same 120 questions about 15 products a second time. In 23 of 120 pairs (19%) the answer moved between fully right and not. The entry price was the same in 79 of 113 pairs (70%). Overall accuracy for those 15 products was similar in both runs: 53 and 48 fully right answers out of 120.

Even with the vendor’s pricing page among its sources, more than a third of answers got a price wrong.

Study 004, Finding 4

05 / 08What it means for vendors

Several of the causes we could trace sit on pages the vendor controls.

1. Retire old prices everywhere you printed them.A blog post and two other pages on vendors’ own sites still carried old prices, and answers repeated them.
2. Make the rupee price readable without a visit from India.A pricing page that switches currency by country shows dollars to a reader outside India. Answers about such vendors in this study quoted dollars or said no rupee price exists. A dedicated India pricing address, with prices in the page text, leaves less to chance.
3. Put every billing option in plain text.Most pages in this study showed one billing term by default and hid the others behind a toggle. A table with each plan and each term leaves less to guess.
4. When you rename a plan, say what it replaced.Plan names the vendor no longer shows, with prices attached, were still in answers: 58 prices in the main run.
5. Ask the engines, then ask again.On a second asking, one answer in five moved between fully right and not. A single screenshot says little about what buyers see.

06 / 08Limitations

• One day, one market. All answers were collected on 5 October 2026 from India. Engines and prices both change; a different week could give different numbers.
• Thirty products, chosen to be checkable. We kept only products whose official page shows a rupee plan list. The results say nothing about vendors that do not publish prices.
• A strict test of the whole answer. An answer with ten prices and one stale figure counts as not fully right. We report the price-by-price and entry-price results alongside for that reason.
• A generous test of each price. We allowed 5% either way, accepted any billing basis, and counted struck-through list prices as current. 28 prices matched only a struck-through figure.
• Outdated is undercounted. Many pricing pages load prices by script or by country, so archived copies often show no rupee prices. Where we could not find an earlier official price, a stale figure is counted as wrong.
• Some views were not captured. A few billing views could not be read (for example Microsoft 365 plans without Teams). Prices for those were set aside as unverifiable, 45 in all.
• We observe what was cited, not why. The citation analysis is an association; it does not show what the engine actually read.

07 / 08Data and reproducibility

The price book, the graded prices, the answer-level verdicts and the citations are published under CC BY 4.0. Our analysis code re-creates every number on this page from the raw answers and the coding files. Version 1.0 corresponds to this publication; corrections will be issued as new versions with a changelog.

• ground-truth.csv274 official prices, with source page and dateCSV
• answers.csv360 answers, one row each, with verdictCSV
• prices.csvEvery graded price, with its matchCSV
• citations.csvEvery cited URL, with source typeCSV
• history.csvEarlier official prices and where we found themCSV

For the full answer texts, questions about the data, or a reanalysis you would like us to check, write to research@citationresearch.com.

@techreport{citationresearch2026study004,
  author      = {{Citation Research}},
  title       = {The price is wrong: how often AI answers quote the current price of software in India},
  institution = {Citation Research},
  number      = {Study 004},
  year        = {2026},
  url         = {https://citationresearch.com/research/ai-software-pricing-accuracy-india}
}