Methodology & sources
This page explains how we set each label, what counts as a source, and how the announced-vs.-found scoreboard is built. A plain-English summary comes first. The strict rules that follow it are for anyone who wants to check our work. The same rules are in the project’s process documents.
How this works
Food Robot Index is a free list of robots working in restaurants and other food businesses. Every entry needs a public source, like a company announcement, a news story, or an official filing. We write down the date and keep a copy of the page. We don’t use private tips, and we never guess. If the proof is thin, we mark the entry “not yet confirmed” instead of leaving it out. When the news goes quiet, we ask for a fresh check. We don’t assume a robot is gone. This is a public record, not investment advice.
What the labels mean
- in use
- A dated source says the robot is working at this named place, in everyday service.
- in a trial
- A dated source says the robot is at this named place, but as a test or trial. It is not in everyday service yet.
- announced
- The company, or a major news outlet, said a robot will be installed at this named place. We have no sign yet that it is there.
- removed
- A robot was here before, and a dated source shows it is gone. A closed restaurant counts. We never say the robot “failed.”
- not yet confirmed
- We saw a public sign, such as a photo or a job post, but not enough proof to call it in use, in a trial, or announced.
The full rules, written for people who want to check our work, are on the methodology page.
Related documents on this site: verification rules (markdown), source monitoring plan (markdown).
Public sources only. Food Robot Index uses no logins, no scraping that breaks a site’s terms or robots.txt, and no private or leaked data. Every install needs a source. If the evidence is weak, use unconfirmed. Do not omit the signal, and do not overstate it. This is a public record, not investment advice.
Status bars
Each location-level record has exactly one current status. Announcement-level count claims belong in data/announcements.json, not here.
unconfirmed
Use when a public signal exists but does not meet the bar for announced, piloting, live, or removed.
Typical cases:
- a single undated photo
- a user social post that does not clearly name the site
- a job posting that mentions a robot and a city but not a store
- a maker page that claims “deployed with Brand X” and names no location
unconfirmed is a full status. Unconfirmed rows are public, like the others. Confidence is usually low.
announced
The location itself must be identified.
A first-party maker or operator source, or a tier-1/tier-2 news source, states a plan or commitment to install a named robot at that named site. There is no evidence yet that the unit is on site in a pilot or in service.
Chain-wide or count claims (“100 units,” “rollout across the brand”) are not location-level announced rows. Record them as announcements. Create an announced install only when the specific site is named.
A trade-show booth, lab demo, or “coming soon” teaser with no site name is not an install.
piloting
A dated first-party source, or a dated tier-1/tier-2 press or local-news source, says the robot is on site at the named location. The source may call it a test, pilot, limited trial, or “in a restaurant for testing,” but must not describe it as regular production service.
One such source is enough if it is first-party or tier-1/tier-2 and names the site. Otherwise two independent public signals are required.
live
All of the following:
1. The specific location is named (store, venue, or equivalent).
2. The robot product (or a clearly identified product family) is named or unambiguously pictured in context.
3. A dated source says the unit is in regular service, not only on display or in a trial.
Evidence options (either one):
- One dated first-party source (operator or maker newsroom, IR, or SEC filing) or one dated credible press / local-news source that names the location and states it is operating in service; or
- Two independent public signals that together name the location and show ongoing operation (for example local news plus a dated on-site photo; an earnings remark plus local TV). Two reprints of the same press release are not independent.
Trade-show, investor-day, and lab demonstrations are not live restaurant installs.
removed
The site previously had a piloting or live (or, rarely, announced) record, and dated evidence shows the unit is no longer there.
Acceptable evidence is any one of these: a first-party statement that the unit was removed or decommissioned; credible press saying the same; or two independent recent on-site signals of absence plus the prior install record. If a store closed, mark the row removed and note the closure. Do not say the robot “failed.”
Do not mark removed only because sources went quiet. Quiet sources trigger needs_reverification and leave status unchanged until there is positive evidence.
Source tiers
| Tier | Types | What it can do |
|---|---|---|
| 1 | First-party operator or maker newsroom; official IR; SEC filing; on-the-record earnings remark posted on IR or EDGAR | Can establish any status if it names the needed facts and is dated |
| 2 | Established trade press, local news, national news, or a press wire quoting a named official | Same as tier 1 for location-level facts those outlets actually state |
| 3 | Official social accounts of the maker or operator | Can support announced or piloting when they name a site and date; alone they are not enough for live unless they are clearly an official newsroom equivalent (link the underlying post and archive it) |
| 4 | User social posts, reviews, job postings, unlabeled photos or videos | Leads and supporting signals only. A tier-4 item may be one of two independent signals; it cannot be the sole basis for live or removed |
Independence means different organizations, not a syndicated copy of the same release. A maker press release and a wire reprint of that release count as one signal.
Staleness
- For
liveandpiloting, iflast_confirmedis older than 12 months, setneeds_reverification=true. Do not auto-changestatus. - After a failed re-check (no confirming public signal found), leave the status in place, keep the flag, and note the attempt and date in
notesor the weekly change log. A second annual miss still does not equalremoved. - For location-level
announced, review if the stated target date is 6 months past, or if 18 months have passed with no target date and no on-site evidence. The record may stayannouncedor move tounconfirmed; do not invent a removal. - For announcements (count claims), use the same 6-month-past-target / 18-month review. Possible announcement statuses:
active,revised,withdrawn,fulfilled,expired.expiredmeans the target window passed without support for fulfillment. It does not mean the claim was false when made.
Proposed windows are also listed in open-questions.md and can be tightened after the first seed pass.
Social posts, reviews, and job postings
- Official maker/operator accounts: treat as tier 3. Archive the post. Quote only the fact-bearing line.
- User posts and reviews: usable only if they are public, dated or date-inferable, and identify the venue. They support other evidence. They are not a standalone
livesource. - Job postings (“robot technician,” a model name, a city or store): investigation leads. They may justify an
unconfirmedrow if they clearly refer to a robot already associated with a site; they do not by themselves make a sitelive. - Do not use private groups, deleted posts that were not archived, or anything behind a login.
Archiving
When a source is used to add a record or change a status, capture a public archive, preferably the Wayback Machine (https://web.archive.org/save/{url} or a save that produces a snapshot URL). Store that URL in archive_url.
If the archive fails, keep url, accessed_date, and a short quote_excerpt, set archive_url to null, and note the failure. Do not use an archive service that requires a login.
Re-archive if the live page changes in a way that affects the cited fact.
Removals and conflicts
- Prefer first-party or press evidence of removal.
- If sources conflict (one says live, a later one says removed), record both, set status to the later well-sourced fact, lower
confidenceif the later source is weaker, and explain innotes. - A maker “customer logo” page with no date and no site is not evidence that every logo still has a live unit.
Confidence
high: tier 1 or 2, dated, names site and product, and is not contradicted.medium: meets the status bar but is thinner (older, official social, or two weaker signals).low:unconfirmed, or a status that rests on the minimum bar and has unresolved questions.
Map and timeline
- The timeline slider shows the data as it stood on a date. A row appears on its
first_seendate and takes the status that itsstatus_historygives for that date. - Dates show when an install was first publicly reported. They do not always show the day it switched on. Coverage is also denser in recent years.
- If every source for a row is undated,
first_seenis the day we first saw the page, andfirst_seen_basisisrecorded. The map shows the row from that date. The bar chart leaves it out. - Removed sites stay on the map as removed after their removal date. A checkbox hides them.
- Many sources name a city or state but give no coordinates. For those rows we look up the middle of the city (or state) on OpenStreetMap and sets
geo_precisiontocityorregion. The pin has a dashed ring, and the detail panel says "Approximate (city center)". The pin does not mark the site. A row that names no place has no pin.
Browse, search and near me
- Browse pages group the same records by kind of robot and by country. They add no records and no claims. Counts on them come from the same data as every other page.
- Search runs in your browser against a small index of chain, maker, city, country and robot-kind names that is built with the site. Nothing you type is sent anywhere.
- Near me works out distances in your browser. It asks your browser for your location only after you press the button, and your position is not sent to us or to anyone else. A typed ZIP code is placed using a table of 3-digit ZIP prefix centers worked out from GeoNames postal code data (CC BY 4.0). Most places are pinned to a city, region or venue, not a street address, so distances are a rough guide.
AI-assisted summaries
- Some chain and maker pages open with a short summary. It is labeled “Summary written with AI help from the records below, checked against the records by a script.”
- An AI model is given only that page’s own records as data, with no web access. A script then checks every number, date, place, name and status word in the text against those records, and rejects text that leaves out the company’s number or the number found by name, or that does not say the two measure different things.
- The script checks facts, not fairness or judgment. The approved text is saved and the site build never asks the AI again. If the records change so a number no longer matches, the summary is hidden until it is redone.
- Pages with only a few records get no summary. The records and sources on the page are always the evidence. A summary is never a source.
What counts as a robot
We list robots that make, cook, serve or deliver food and drinks inside a venue. That includes in-restaurant runner and bussing robots, fry stations, bowl lines and drink makers. As of Nov 18, 2026 it also includes every automated coffee kiosk, such as the Touch Coffee Smart Bar: a self-serve machine that makes a drink with no staff at the counter. A kiosk is listed only when a public, dated source names the place. A kiosk that only shows up in a company’s own social post is labeled “not yet confirmed”. Sidewalk delivery robots, cleaning-only robots and factory lines stay out.
Out of bounds
- No record without a public source.
- No credentials, email harvesting, or off-the-record tips used as the source of record. A tip is a lead. It becomes usable only when a public source is found.
- No ToS-violating scraping. Prefer official RSS and manual reading of public pages. Check
robots.txtbefore any automated fetch. - Paywalled articles: cite them only if a public headline/lede or archived public snapshot supports the fact, and look for a free corroborating source. Do not create accounts to bypass paywalls.
- Do not guess street addresses or precise coordinates. See
data/schema.mdforgeo_precision.