Veritaset · Right Deal Index · Mobile plans, Europe and USOctober 2026 benchmark
Right Deal Index · October 2026
Don't take your AI's word for it
We asked ChatGPT, Gemini and Grok 24 everyday questions about phone plans in Europe and the US, and checked every answer against the operators' own websites. They found the cheapest plan that fits about half the time in Europe and rarely in the US.
35–48%
of app answers named the cheapest plan that fits
+37–47%
average extra cost when the plan fits
79–92%
of Veritaset's answers right on new, pre-registered questions
Summary
The apps get it right less than half the time. On 21 questions with an operator-checked answer, ChatGPT named the cheapest plan that fits in 42% of scored answers, Grok in 48% and Gemini in 35%: 46–60% in Europe, 0–33% in the US. The intervals overlap, so this doesn't rank the apps.
The mistakes cost money. When an app named a plan that fits, it cost on average 37–47% more than the cheapest one. 10 of 60 scored answers cost more than double.
Most misses: a cheaper plan was available. In 22 of 35 wrong answers a cheaper plan that fits existed and the app didn't recommend it.
Verified data fixes most of them. Shown a verified alternative and asked neutrally to check both, the apps switched to the right plan in 27 of 35 wrong answers.
We hold ourselves to the same test. Graded by the same answer key, Veritaset was right on 19 of the 21 questions. On new questions written after the test, it was right 92% of the time in the US and 79% in Europe.
Part 1
Which assistant gets you the right deal
An answer counts as right when the plan fits the stated need (data, minutes, contract, eSIM, travel) and costs at most about €1 a month more than the cheapest plan that fits, at the operator's own price. We used the apps as people do: a new chat per question, memory off, no connector. Three of the 24 questions are excluded for everyone, because the cheapest offer on the test day can't be verified.
Exhibit 1 · Consumer apps
All three apps name the cheapest plan that fits in fewer than half of the questions
Right-deal rate on the same 21 questions, with 95% interval. Scored answers: ChatGPT 19, Grok 21, Gemini 20, Veritaset 21.
Consumer appVeritaset
October 2026 benchmark: ChatGPT (Pro, reasoning "medium"), Gemini (Gemini Pro), Grok (Auto, standard subscription); 16 European markets and the US. Europe: ChatGPT 6 of 13, Grok 9 of 15, Gemini 7 of 14. US: ChatGPT 2, Grok 1, Gemini 0 of 6. Veritaset: its answer on the day of the test.
Exhibit 2 · The cost of mistakes
The plan an app recommends costs 37–47% more on average
Answers by extra cost over 12 months compared with the cheapest plan that fits. Average among answers with a valid plan: ChatGPT +43%, Grok +37%, Gemini +47%.
ChatGPTGrokGemini
Largest overpayment: ChatGPT's contract plan in Warsaw, about €82 a year against €23.50 (+249%). Where an app's own stated price already lost (17 answers), we used its figure, so these are lower bounds. "No valid offer": doesn't fit the need, isn't sold, or needs a port-in from certain operators. Three answers based on comparison-site deals are not shown.
Exhibit 3 · Developer tools
The command-line versions were no more reliable
Right-deal rate on 58 European questions, own web search only, two runs per question, with 95% interval.
October 2026 benchmark. In the US (39 questions) Claude Code was right in 13% and Grok in 53%. Different questions from the apps; an answer we couldn't check counts as not right.
Part 2
Why assistants get it wrong
To give the right deal, an assistant has to see the whole market, use today's offers, read each offer's details and check its conditions. We classified each of the 35 wrong answers by the first step it failed.
Step 1
22 of 35
A cheaper plan was available
Often a smaller brand, a newer offer or an annual prepaid deal.
US, 15 GB with eSIM: all three apps picked Tello at $240–280 a year. Infimobile's annual plan costs $108.
Step 2
1 of 35
Not on sale
A plan the operator doesn't sell (any more).
Vienna: "spusu 4.000", which spusu doesn't sell.
Step 3
2 of 35
Wrong details
The right operator, but the wrong price, allowance or fees.
Vienna: Lycamobile National XS presented as 3 GB; it is 2 GB.
Step 4
10 of 35
Conditions not met
Data slowed to a few Mbit/s, two separate packs, or a price that needs a port-in.
Milan: WindTre GO Unlimited, sold only to customers porting from listed operators.
In two cases (Helsinki, Budapest) the cheapest plan was already in ChatGPT's own search results and it recommended another: the search worked, the comparison didn't.
Part 3
What verified data changes
After the test we went back into the original conversation of every wrong answer. We named the answer key's plan with its operator page and check date, and asked the app to check both plans on the operators' websites and say which is the cheapest that fits. The prompt was the same every time, didn't say which plan was right, and the protocol was written before the first run.
Exhibit 4 · Shown a verified alternative
The apps switched to the right plan in 27 of 35 wrong answers
Wrong answers followed up in the original conversation, by the step the first answer failed.
First answer failed at
Wrong answers
Switched to right plan
Kept or other wrong plan
No answer
1. A cheaper plan was available
22
20
1
1
2–3. Not on sale, wrong details
3
3
0
0
4. Conditions not met
10
4
6
0
All
35
27
7
1
Scored by code with the same key. By app: ChatGPT 7 of 11, Grok 9 of 11, Gemini 11 of 13. ChatGPT ran on a Plus plan for the follow-up (Pro in the test).
When the app had missed a cheaper plan, a dated, sourced alternative fixed it almost every time. When it had misjudged a condition, it mostly kept its pick. A verified layer has to state conditions as clearly as prices. One limit: the alternative we named was always the right one, so this shows apps accept a correct suggestion after checking it, not that they would reject a wrong one.
Our own error rate is part of the result
Veritaset is graded by the same answer key. On the 21 scored questions its answer on the test day was right in 19. In Madrid it recommended simyo's €3 line, which needs a main simyo contract; in Berlin its plan cost €12.03 a year more than Lebara's, just over the margin. Among the excluded questions, its rural Vermont pick doesn't fit (70 GB at full speed, not unlimited); counting it, 19 of 22.
Because we used these questions to improve our data, we also wrote new questions and fixed them before any result. On those, Veritaset's data from the time of the test was right on 23 of 25 US questions (92%, interval 80–100%) and 44 of 56 European ones (79%, 68–89%). A random audit in October found 96.9% of 1,690 mobile-plan facts correct (interval 96.0–97.6%).
Accurate consumer information is hard, for an agent and for us. People need an independent check on what their agent tells them, with no commission on the answer.
Free · independent · no commission
Fact-check your AI
Add Veritaset to your assistant. Every price comes from the operator's own website, with its source and check date. Mobile plans in 34 markets, home internet in 34, household energy in 4.
In your product. Shopping and assistant builders can call the same connector or ask for API access: data@veritaset.com.
Independent. No affiliate links, no commissions, no paid placement, no commercial relationship with any operator. Ranking rules: veritaset.com/method/.
Method, limits and how to reproduce it
Method
What it measures. How often AI assistants find the cheapest plan that fits, and what their mistakes cost.
Questions. Everyday needs: data, minutes, lines, contract, eSIM, budget, network, travel. Consumer apps: 24 questions (16 European markets, 8 US). Developer tools: 58 European and 39 US questions, two runs each, on 20–21 October 2026. Held-out round: 83 new questions (58 Europe, 25 US), files hashed and pre-registered before any run.
Consumer apps. ChatGPT on Pro with reasoning "medium"; Gemini with Gemini Pro; Grok in Auto mode on a standard subscription; all on 22 October 2026 (the follow-up ran on 2 November 2026), one new chat per question, memory and chat history off, no connector. Free plans may use smaller models and search less. Answers read from share links and scored by code.
Answer key. Every plan any assistant or Veritaset recommended was checked on the operator's own website, with a quote per fact, except 17 answers whose own stated price already lost. Sites that refuse automated reading were checked by hand in a normal browser. For two US questions the reference price comes from our own record, re-checked on Infimobile's site, infimobile.com, on 2 November. Offers an app found that we lacked were added to the key, which can only lower our score.
Scoring. By code: right if it fits and costs at most about €12 a year more than the reference. Only a price any new customer can get from the operator counts; a code to type, a port-in from certain operators, a membership or a second product doesn't fit a general need. Non-euro prices at fixed reference rates. A plan paid 12 months up front without further commitment counts as no contract, at its first-year price. Infimobile's prices exclude taxes.
Exclusions. Lyon (a Lebara offer in its page data on 27 October we couldn't see on 22 October or 2 November), El Paso (Mexico passes we can't read on the operator's site) and rural Vermont (a Visible offer that ended on 29 October). Key corrections on 2 November: simyo's €3 line in Spain is an extra line, and US Mobile's Unlimited Starter has 70 GB of full-speed data on Verizon. Key corrections made during the follow-up were applied to the October answers and follow-ups alike.
Limits
One conversation per question per app. Apps vary between runs; 21 questions show the size of the gap, not a ranking.
All apps were used from one account in Europe with Dutch as the interface language; location and language can change results.
The answer key covers plans that assistants or Veritaset surfaced, not every plan in the market. A cheaper plan nobody named would not show up as a miss.
Deals found only through intermediaries, such as a comparison site's voucher, are outside the operator-price rule (three answers, UK and Spain).
Travel eSIMs for trips abroad are not yet in our data.
The accuracy audit's checker is a model of the same family as the one that reads the pages.
Veritaset publishes the data used in the comparison, so we have an interest in the result. That is why we report our own error rate.
Reproduce it
The 24 questions and exact prompt, a share link for every consumer-app conversation, the follow-up prompts and answers, the answer key with the operator quote behind each fact, and the score and failure step of every answer, with checksums and a list of corrections: veritaset.com/benchmark/2026-10, under CC BY-NC 4.0 (free to reuse with credit, not commercially). Also on Zenodo (DOI 10.5281/zenodo.23941830) and Hugging Face. Veritaset's price database itself is not part of it. The index will be published per category, market and month.