A prototype client report by The AI Pipe, 26 September 2026
Italian buying questions: what an AI visibility report can actually claim
Asked the same Italian-language questions, the two surfaces point to different places. Counting each cited domain once per answer, OpenAI's 22 answers that searched the web pointed to a brand's own site 44 times out of 56; Perplexity's 30, 2 times out of 193, citing publishers and shops instead.
Home coffee machines, an illustrative category: no client relationship, and Cognitive is not measured. Ten Italian questions copied from Quora and a consumer forum, asked on two programmatic surfaces, three runs each: 60 answers, 0 failed. Italian-language observation, location uncontrolled.
The report a brand would receive
Brands named in five answers or more. Select a cell to read the answer behind it.
OpenAI API with web search
Own site in 44 of 56 cited domains. Answered without searching: 8 of 30 (hatched).
Swipe the plan sideways to see questions 1 to 10.
Perplexity Sonar
Own site in 2 of 193 cited domains. Searched in all 30 answers.
Swipe the plan sideways to see questions 1 to 10.
Perplexity Sonar, question 4, run 1, asked at 15:23:14 Rome time, 1 web search
Qual è la migliore macchina da caffè sotto i 300 Euro?
- De'Longhi named
- yes
- recommended
- yes
- own site cited
- no
- pages cited
- 7
Reading based on: “Migliore scelta “tuttofare” sotto 300 €: De’Longhi Magnifica S. - Migliore per espresso manuale e budget contenuto: De’Longhi Dedica EC685 oppure Stilosa.”
Cited pages
- femmeactuelle.fr, publisher or guide
- casaprova.it, publisher or guide
- tomshw.it, publisher or guide
- libero.it, publisher or guide
- aranzulla.it, publisher or guide
- market.com, publisher or guide
- trovaprezzi.it, shop or price listing
7 more domains consulted by the search but not shown as citations.
Counts per run, never averaged
Each figure lists runs 1, 2 and 3, out of 10 answers per run. Three runs of this fixed panel show observed variation, not a margin of error.
| Brand | OpenAI | Perplexity | ||||
|---|---|---|---|---|---|---|
| named | recommended | own site cited | named | recommended | own site cited | |
| De'Longhi | 4, 4, 4 | 3, 3, 4 | 4, 3, 3 | 5, 6, 4 | 3, 6, 3 | 0, 0, 0 |
| Nespresso | 4, 4, 5 | 4, 4, 4 | 3, 3, 4 | 5, 5, 4 | 2, 4, 2 | 0, 0, 0 |
| Lavazza | 4, 3, 5 | 2, 2, 4 | 3, 3, 4 | 4, 2, 4 | 2, 2, 3 | 0, 0, 1 |
| Dolce Gusto | 1, 1, 2 | 1, 1, 1 | 0, 0, 0 | 2, 1, 3 | 2, 1, 1 | 0, 0, 0 |
| Illy | 2, 1, 1 | 2, 1, 1 | 2, 1, 1 | 3, 1, 2 | 1, 1, 1 | 0, 0, 1 |
Where the citations point
Cited domains per run, each counted once per answer, by the kind of site.
OpenAI API with web search
16 brand, 1 shop, 3 publisher, 0 forum; 14 brand, 2 shop, 4 publisher, 0 forum; 14 brand, 1 shop, 1 publisher, 0 forum. Consulted but not shown: 16, 34, 13 domains.
Perplexity Sonar
0 brand, 23 shop, 43 publisher, 2 forum; 0 brand, 20 shop, 41 publisher, 0 forum; 2 brand, 19 shop, 42 publisher, 1 forum. Consulted but not shown: 58, 68, 64 domains.
One statement, followed to its sources
Question 4, “Qual è la migliore macchina da caffè sotto i 300 Euro?” In all three runs Perplexity recommends the De'Longhi Magnifica S, and never cites delonghi.com. The case was picked by a fixed rule: of the statements a cited page contradicts, the one repeated in the most runs.
…se cerchi un’automatica con macinacaffè integrato, la De'Longhi Magnifica S è la candidata più forte nei risultati: viene indicata come “la migliore macchina da caffè automatica a meno di 300 euro” e come best-seller affidabile.[1][2]Counted: De'Longhi named in 3 of 3 runs, recommended in 3, own site cited in 0.
“De'Longhi Magnifica S : la meilleure machine à café automatique à moins de 300 euros”Selection updated 18 September 2026 (the page says so; its metadata says the same).
Supports the words in quotation marks, translated. It is a French selection, which the answer does not say.
“Se preferite un'automatica De'Longhi, la Magnifica S a €399,00 ha 4,3/5 su 1.853 recensioni.”
“Philips 3200 LatteGo - automatica con latte sotto 300€”Published 29 April 2026, modified 26 August 2026 (page metadata); our copy was downloaded six minutes after the answers.
Contradicts the Italian price point: it lists the machine at €399,00 and names another machine as its automatic under €300.
What a report can claim from this: the answer's quoted verdict exists, but on a French page, and the Italian page cited beside it prices the machine above the threshold (a contradiction our reader flagged in 3 of 3 runs). The reader model did not flag the French source; the builder agent's own review did (no Italian speaker was involved). What it cannot claim: that the Magnifica S never sells under €300 in Italy, or why the engine chose these pages. OpenAI, on the same question, picked the Dedica Duo, the Sage Bambino, the Dedica Duo in runs 1, 2 and 3, citing delonghi.com for the Dedica Duo in runs 1 and 3, and two buying guides (novebar.com, brewmance.fr) for the Bambino in run 2.
What the brand could test
- Observed
- In all three runs the answer calls the Magnifica S the best machine under €300. The only pages attached to that statement are a French selection and an Italian guide that prices it at €399,00. No De'Longhi page is cited, and no cited page shows an Italian price under €300 for this machine.
- Hypothesis, not a finding
- The engine retrieves no brand page and no Italian page that puts this machine under €300, so the under-€300 verdict comes from the French selection. A dated Italian price on the brand's own product page, or on a retailer listing the engine already cites, might change what the answer rests on.
- Test
- Freeze this question and its configuration and run it on a schedule for several weeks before and after the change. Count the runs where the Magnifica S statement cites a De'Longhi page, and the runs where it cites an Italian page with an Italian price under €300. Today both counts are 0 of 3. Only a rise larger than the before-period variation would count, and even then it shows association, not that the page caused the citation.
What to build, what to buy, what to keep manual
Trackers on the market already separate brand and source visibility by model and country (Peec advertises this). The difference a report can add is the evidence behind each line and the decision it supports. Two products could come out of this sample.
A client report, reviewed by a person
Build this first.
- Own
- Frozen question panels per category, the configuration register, the four counting fields, the evidence store (answer, copy of each cited page, its dates), the source table, the report template.
- Buy
- API access to each answer surface under terms that allow storing and reporting, and a page renderer for the 18 of 118 cited pages a plain download could not read.
- Keep manual
- Choosing questions with a native Italian reviewer, borderline readings, the case shown to the client, and any consumer-app observation (by hand, few, labelled).
- This sample shows
- It runs end to end on real Italian questions for $1.61 of API calls, and its automated readings held when the builder agent re-read a sample (16 of 16 recommendation readings, 12 of 12 source checks); no Italian speaker has checked them.
A recurring monitoring product
Not yet.
- Own
- The same rules plus scheduling, alerts and a client dashboard.
- Buy
- Collection at scale with rights to archive and resell answers: not established for Perplexity through a reseller, not available for Gemini grounded on Google Search under its terms, and excluded for consumer apps.
- Keep manual
- Still the readings and source checks, until their error rate is measured on larger samples.
- This sample does not show
- Stability over days (our three runs were minutes apart), any effect of location, market share, or that the reader's accuracy holds at volume.
Method
- Surfaces and configurations, fixed before collection. OpenAI API with web search: OpenAI Responses API, model
chat-latest(returned as chat-latest), web_search tool with tool choice left to the model, up to 4 searches, 3000 output tokens, approximate location Italy, Milano. Perplexity Sonar: modelperplexity/sonarthrough OpenRouter, search context low, country IT sent. No system prompt on either. The question text is sent exactly as posted. - Collection window, Rome time: OpenAI 15:23:11 to 15:24:33, Perplexity 15:23:11 to 15:24:03, 26 September 2026. Runs 1, 2 and 3 were sent in sequence within that window, so they measure short-term repeatability only.
- Location: an Italian location was requested on both surfaces; its effect was not tested on this panel. Results are an Italian-language observation with location uncontrolled.
- Four fields per brand and answer, never merged: named in the prose (link labels do not count); recommended (a reader model proposes a verdict with the exact words, code rejects any quote not found in the answer); own site cited (a citation to a domain the brand owns); support (for every statement naming a brand that carries a citation, the stored copy of the page is read: supported, partial, contradicted or not assessable, again with verbatim page words checked by code).
- Absences defined before counting: failed collection (0); answered without searching (8 OpenAI, 0 Perplexity); consulted but not shown is never a citation; one page and one domain count once per answer; a page we could not read makes its statements not assessable, never unsupported.
- Brand aliases (De'Longhi, De’Longhi, DeLonghi, De Longhi), co-branded models counted for both brands, model names alone not counted, ordinary-word names counted only when capitalised. 21 brands were named at least once.
- Dates are kept apart: 100 of 118 cited pages were read; 47 declare a publication date and 47 a modification date; every page keeps its retrieval time. No single “freshness” value is computed.
- Tests (vitest) cover aliases, repeated URLs, consulted against cited, negative mentions, missing citations, failed collections, statement extraction and this case against the stored pages.
The ten questions and where they come from
Copied, not written: seven Quora question titles and three first posts from the Tom's Hardware Italia forum. Search suggestions and “People also ask” boxes were not used, for lack of a permitted way to collect them. No Italian speaker reviewed the panel; the posts are not a representative sample of what Italians ask assistants.
- 1Qual è la migliore macchina da caffè a capsule?Quora (Italian), question title, posting date not shown; edits: none
- 2Qual è la migliore macchina da caffè espresso per casa?Quora (Italian), question title, posting date not shown; edits: none
- 3È meglio la macchina da caffè Lavazza o Nespresso?Quora (Italian), question title, posting date not shown; edits: none
- 4Qual è la migliore macchina da caffè sotto i 300 Euro?Quora (Italian), question title, posting date not shown; edits: none
- 5Conviene comprare una macchina del caffè di marca Dolcegusto, Nespresso o Lavazza?Quora (Italian), question title (unanswered on Quora), posting date not shown; edits: none
- 6Meglio una macchina per caffè a cialde o a capsule?Quora (Italian), question title (unanswered on Quora), posting date not shown; edits: none
- 7Quale marca di macchinetta per i caffe con cialde mi consigliate ?Quora (Italian), question title, posting date not shown; edits: none (spelling and spacing kept as posted)
- 8Ciao a tutti devo comprare una nuova macchina per il caffè ma sono molto indeciso su quale prendere; in particolare non so se comprare una a cialde o in polvere (tipo bar). Ne prendo uno al giorno quindi l'utilizzo non è eccessivo, ho un budget di €200. Cosa potete consigliarmi?Tom's Hardware Italia forum, first post of the thread 'Nuova macchina del caffè', posted 2022-10-03; edits: none (whole first post)
- 9Ciao, vorrei fare un regalo ai miei genitori e sono ricaduto su una nuova macchina del caffè. Su quale macchina a cialde o a capsule compatibili potrei optare? Budget: 80/100 euro circa.Tom's Hardware Italia forum, first post of the thread 'Consiglio acquisto macchina caffè', posted 2024-12-08; edits: closing line 'Grazie in anticipo.' removed
- 10Buonasera, mi aiutate? Attualmente uso una macchina per cialde ma il caffè non viene come vorrei pur usando le cialde Illy. Sto pensando di passare ad una macchina che macini all'istante i chicchi che ne pensate ?Tom's Hardware Italia forum, first post of the thread 'Consiglio per acquisto macchina per espresso automatica', posted 2026-02-08; edits: none (whole first post)
Cost register
| Pilot, question 1 on both surfaces, before the panel | $0.086 |
| OpenAI, 30 answers | $1.358 ($0.045 per answer) |
| Perplexity, 30 answers | $0.166 ($0.0055 per answer) |
| Total API spend, against a hard cap checked before every request | $1.61 of $10.00; 0 retries, 0 failures |
| Downloading the cited pages | no charge; 118 pages, 18 unreadable |
| Reader passes (recommendations, source checks) | run on a flat-rate model subscription, not metered here |
| Review | 16 readings, 12 source checks and the case re-read by the builder agent (Claude), not by a person or an Italian speaker; time not recorded |
| Italian-language review | none; required before a client sees a report |
The API is the smallest line. What a usable client report costs is mostly the review, which this sample did not time.
Rights register
| OpenAI API with web search | Used under the API terms. Answers are stored and quoted here with their citations shown as links. |
| Perplexity Sonar through OpenRouter | Used through a reseller. Perplexity's own API agreement was not reviewed, so archiving and reselling these answers is not established. |
| Gemini with Google Search grounding | Not used. Its terms restrict analysing, extracting and storing grounded results, which rules it out for a monitoring database without written permission. |
| ChatGPT app, Google AI Overviews | Not collected. Consumer terms and Google's crawling rules exclude automated collection; no manual observation was made for this sample. |
| Cited pages | Downloaded once to check statements; not republished. The page shows short quotes only. |
| Questions | Short question texts quoted with a link to their public source. |
Our reading of published terms, not legal advice.
Readings and checks
- Recommendation readings: 155 kept, 0 rejected by the quote check. Re-read by the builder agent (no Italian speaker), 16 drawn at random: 16 of 16 agree (one borderline: a comparative “conviene di più se” counted as a conditional recommendation).
- Statements naming a brand with a citation: 395 brand, statement and page triples, 283 statement and page pairs per answer. OpenAI 44: 14 supported, 22 partial, 0 contradicted, 8 not assessable. Perplexity 239: 76 supported, 117 partial, 9 contradicted, 37 not assessable. Of the not assessable, 29 are pages we could not read (blocked, gone, built by JavaScript) and 16 are pages that say nothing on the point.
- “Partial” is broad: most are pages that list a product without the ranking or the attribute the answer gives it.
Archive
Every answer with its configuration, time, run, status and citations; the readings; the source checks; the source table; the panel: calls.json, readings.json, support.json, domains.json, panel.json.