Opinion Meter
The product and the repository (samco26/Opinion-Meter) are called Opinion Meter. The earlier search site, What People Think, is a separate project whose engine this one salvages.
A browser extension that shows, beside anything you are about to click, what people actually think of it: a small bar, how many voices, one sentence — and the full picture on click.
Current behaviour — 23 September 2026, v1.0.4
v1.0.0 (the day's 0.17.0 work, declared version one) makes the bar and the card one reading, puts the card on every site, and adds "Analyse this page's subject". Beside every Google result the verdict names its subject ("63% positive opinion of Sony Australia"); the badge's cards fold on a press, with no ×s. One reading: the bar's quick reading is the reading. The server keeps the bar's sample (the entries and each one's verdict) for a quarter of an hour, and a card asked for in that time — the recurring opinions list, the prefetched card — is built from that very sample with the verdicts locked, so its numbers are the bar's to the digit; the list no longer pushes numbers back into the bar, on the page or in memory. A card asked for later (the sample aged out, or no bar came first) is read afresh and becomes the subject's reading for the next search; a bar already on a page keeps what it shows. X is read for the bar and the card alike, once per subject. The card on every site: the badge at the top right now appears on every site, not only those reached from Google: the site's reading is fetched by the site's address (the domain, never the page's path) and the name it declares for itself, and shared with the site's other tabs; the × hides it on that page until the page is next loaded; the menu switch ("Show the card on every site") turns it off everywhere. The second sheet: the badge is two sheets — the site's pill on top and, tucked behind it, a tray whose lower strip is the button "Analyse this page's subject" (centred, a faint green-into-red tint sweeping across it). The pill opens into the site's card on a press; the strip rides down with it. Pressed (after a one-time note), the page's visible text — reviews and comments first, up to about twelve thousand words, never anything typed into forms — goes to the server once; a ring of light circles the sheet while the page is read. The server names the page's subject, then at the same time sweeps the platforms for it and reads the page for the opinions written on it (each copied verbatim and checked against the page), then writes one reading: a bar and figures, a summary, the sources (the page's own favicon among the platform tiles), and the pros and the cons in two columns of green and red bubbles, each opening the words behind it. A review found on the page counts as three platform posts. Fewer than eight opinions in all: "Subject does not possess sufficient sentiment footprint"; a page about nothing in particular: "Cannot determine single subject", centred on the strip (two even lines, the strip growing, when the pill is too narrow for one); a page the server could not read at all is not guessed at — the strip says so and can be pressed again. The tray then grows into the page card beneath the pill (which stretches to the same width); only its pros and cons scroll; the two cards open and close on their own (the site's by its ×, a press on it or scrolling; the page's by its header, its × or a press anywhere on it but the lists; Escape closes the topmost, a press outside both), and an open card tints grey under the pointer wherever a press would fold it. Sending page text and site addresses is a change to what leaves the browser: section 9, the privacy page, the menu and the Firefox declaration say so.
Earlier — 16 September 2026, v0.16.6
v0.16.6 fixes a shared style that accidentally hid the loading sweep until sentiment arrived. The browser check now verifies that loading bars are visible and have size, not only that they exist.
The latest shared extension draws loading bars for every recognised result as soon as the config door permits it, before waiting for sentiment. The main query gauge stays above the results, including searches with a knowledge panel, and keeps an empty gauge when no verdict is available. Expanded cards keep the platform controls together where space permits and the opinion count on a separate line; long text wraps. These current rules supersede the older placement and loading descriptions below. All four browser packages share these fixes.
The drawer starts compact: See recurring opinions between the platform icons and count reveals the scrolling opinions section. Bars stay below the drawer while scrolling; hover gently brightens the bar without scaling or outline rings and fades in a tooltip-coloured background. The floating site card grows for its name and verdict, and its hover paragraph follows dragging. Icons have rounded transparent corners; Safari and Firefox show cream in dark mode and grey in light mode, with cream as the compatibility fallback. All four browser packages are always built and published together unless the owner asks otherwise.
The main card summarises opinion about the search query. A result's bar is about the thing its page represents — a product, a service, a film, an app, a company, a site's front page, a video, an article — as people regard it over time, named by rules where a site's address carries the identity and by one batched AI call otherwise. The query and a result about the same thing share one reading (the youtube.com result and the query "youtube" are both YouTube). A page about nothing in particular (a how-to, a login page) gets no bar. Only without an AI key does an unnamed page fall back to a reading of the page itself, then its website.
Summaries and the bar's sentence describe the subject's general standing, never the news of the moment; a separate "Lately" line appears only when a development of the past two weeks has clearly swung opinion, resting on at least three dated entries, with roughly when it happened. Momentary "is it down?" reports do not count as opinions. The sample the model reads takes turns across platforms and caps any single thread, so one busy platform or one outage thread cannot drown the rest; search windows widen platform by platform.
Result bars are drawn on a layer over the page and pinned right after the site name's text (or after the dots where the name shares its line, under a Shopping tile's seller line, after a source label in an AI answer's panel, or in the top-right corner of a card that names its site at the bottom), with "42% positive" and no opinion count, fade in, adapt to Google's dark theme, and are invisible until a reading exists — an empty outline when the evidence is thin. The query's card mirrors Google's pill buttons on the results pages and sits beside a knowledge panel's title. The drawer is anchored to the page beside its bar, shows the three percentages under the bar, a "Lately" line when there is one, and only its recurring opinions scroll. Four packages come from one build: Chrome (which also installs on Brave, Opera, Vivaldi and Arc), Edge (the same package under its own name, for the Edge store), Firefox (its own manifest), and Safari (the folder in Safari's shape, which a Mac job in CI turns into an unsigned Mac app on request; signing and the App Store need an Apple developer account). On sites opened from Google, "take the bar with you" shows the reading as a frosted card at the top right that can be dragged anywhere and fades while scrolled.
Previously (v0.2, 14 September): result bars measured the exact linked page with a labelled website fallback and never borrowed the query's sentiment; result controls were bare bars with placeholders while loading.
The hands cover organic headings, sponsored results, product tiles and AI Overview references. Result controls are bare bars, with counts and scope on hover/focus; unavailable readings are neutral outlines. The main card has no summary subtext. It sits above the right-hand reference column when one can be recognised, otherwise above the main results. Dynamic results are processed in bounded batches, including repeated destinations and results after the first batch.
The drawer preserves the earlier site's summary, category sentiment score, platform buttons, recurring opinions and linked original excerpts. It is created and fetches the full analysis only on click, with no prefetch or background full-card analysis. Its frosted background blurs the underlying Google page; its height follows the content up to the viewport limit. Original query/result context recovers a lost subject on another server instance. Full cards containing excerpts are not cached; gauges and subject identities last at most 24 hours.
Expanded readings search all available indexed dates on Hacker News and Bluesky. Hacker News requests up to 1,000 matches and reads up to 1,000 comments per reading, including comment-text searches. Bluesky uses its maximum 100 posts per page and follows up to ten cursor pages. Both retain partial results on timeouts and disclose limits; this does not guarantee every historical post. See Bluesky's API contract and Algolia's pagination limit. The AI still classifies a bounded sample (60 for a bar, 250 for a full card); collected counts and analysed counts are distinct.
Production server: https://opinionmeter.vercel.app. The owner's CHATGPT and YOUTUBE environment variables take precedence over OPENAI_API_KEY and YOUTUBE_API_KEY. /api/health reports configuration booleans, never key values. Adding keys requires a new deployment. Existing estimates are recomputed after AI is connected. Shared Upstash memory is still needed for reliable polling and global limits across Vercel instances.
This document is the specification. It is written for a reader with no computer-science background and doubles as the brief for anyone (or any AI assistant) working in this repository. Read it before changing anything.
1. The concept
You are about to click a Google result, buy a thing, watch a film, install an app. The crowd's verdict on it already exists, scattered across Reddit threads, X posts, YouTube comments, Hacker News and Bluesky. Nobody gathers it at the moment you need it.
The extension does. On a Google results page, any result that is about a named thing (a product, a film, an app, a place, a company) or an article people are discussing gets a small bar beside it:
- the split of opinion (positive / mixed / negative), counted from real classifications, never guessed;
- how many opinions were read;
- on hover, one sentence in plain words;
- on click, the drawer: the full card with the summary, the split, the recurring opinions and links to the actual posts.
When nobody is talking about something, nothing is drawn. Silence is part of the product.
Version 1 does Google results only. YouTube pages, news articles and eventually every page follow the same design and add one small "skill" each; see the roadmap.
2. Principles (not negotiable)
- Honest numbers. Every percentage is counted from per-opinion classifications, weighted by reactions on a log scale. Fewer than eight on-topic opinions means no verdict. The count is always visible. A positive result is not the same as strong agreement, and the card says which it is.
- Silence when thin. No bar is better than a misleading one. Utility pages, unknown subjects and thin samples get nothing.
- Anything named is a subject. Products, media, apps, places, companies, articles, topics and people. A person is read on their work and public conduct, never their private life. (Decided 15 September 2026, reversing the 14 September rule that people were off-limits.)
- Nothing stored that isn't ours. The server keeps only derived things — counts, the summary sentence, links — for at most 24 hours. Verbatim excerpts are fetched live when a reader opens the drawer, so a post deleted on its platform disappears from the product at once.
- Official doors only. Every platform is read through its official API within its published limits. No scraping of platform websites. Facebook, Instagram and TikTok offer no such door and are therefore not sources.
- The extension stays dumb. Keys, thinking and memory live on the server. The extension sees the page, asks, and draws. Improving the analysis never requires a store re-review.
- Least privilege. The extension asks the browser only for the sites it needs (Google search pages in v1) and sends the server only what it needs (result titles and addresses from those pages, nothing else). What is sent is stated on the privacy page in the same words.
- Attribution. Every opinion shown links back to the thread it came from, with the platform named.
3. Plain-words glossary
| Word | Meaning here |
|---|
| Server | An always-on computer that every copy of the extension can phone. It holds the keys, does the thinking and keeps the memory. Rented from Vercel; any equivalent host would do. Nobody "visits" it. |
| Door | An address on the server that a program can knock on with a question and get an answer (an API route). |
| Memory | A small database attached to the server holding recent answers so that one answer serves everyone (a cache). |
| Extension | The small app a reader installs from the store. It is the front half; the server is the back half — the way the Google Maps app is small because the map lives on Google's computers. |
| Label | The extension's manifest file: its name, version, which sites it may touch, what it is allowed to do. The store reads this most carefully. |
| Hands | The content script: the part of the extension that runs on a page, reads it and draws the bars. Each surface (Google, later YouTube, articles) is one skill of the hands. |
| Brain | The background service worker: the part that talks to the server and hands answers back. It runs inside the reader's browser, not on a server. |
| Drawer | The panel that opens over part of the results page when a bar is clicked. It frames the server's embed page, so the full card, the platform evidence and the opinions are drawn by the server and never need a store update. The same behaviour on every browser (Safari has no side panel). |
| Subject | The thing an opinion is about, named canonically ("Sony WH-1000XM6", not amazon.com/dp/B0…). |
| Gauge | The quick answer for the bar: split, count, one sentence. |
| Card | The full answer: summary, split, recurring opinions with counts, evidence links. |
| Store | The Chrome Web Store (and later Firefox Add-ons, Safari): the only way ordinary readers can install an extension. |
4. How it works
4.1 Where the pieces live
GitHub (this repository: server/ and extension/)
|
|-- push ---------> Vercel -- runs the server: doors, embed page, memory
| |
| '-- calls --> Reddit, YouTube, Hacker News, Bluesky, OpenAI
|
'-- zip, upload ---> Chrome Web Store -- reviews, then installs into readers' browsers
|
A reader's browser v
+------------------------------------+
| Hands reads Google's results page |
| Brain asks the server <----------+----> Vercel
| Drawer shows the card |
+------------------------------------+
Two halves. The server half updates itself every time code is pushed to GitHub. The extension half only changes when a new zip goes through the store, which is why the extension is kept as small and dumb as possible.
4.2 A Google search, step by step
- The reader searches. Google builds its results page as always; Google's computers are never involved in what follows and never know.
- The label says "on Google search pages, run the hands", so the browser starts them.
- The hands pick out the result links — title and address of each — and the query itself.
- The hands bundle those and pass them to the brain, which sends one message to the server's gauge door: "ten results for sony xm6 review; what do people think of each?"
- The server names the subject behind each result (section 4.3), merges duplicates (six of ten results about the same headphones are one subject), and checks its memory.
- Known readings come straight from memory. New readings get a quick lite gauge. The full card is computed only when clicked.
- The server replies with a short list: result 1 → 72% positive, 340 opinions, one sentence; result 2 → same subject; result 4 → nothing found.
- The hands draw a bar beside each result that has one, inside a protected bubble (Shadow DOM) so Google's own styling cannot disturb it. Hover shows the sentence. Click opens the drawer with the card.
- When Google swaps results without reloading, or the reader moves to page two, the hands notice and repeat from step 3.
Readers see bars about a second after the results for anything already in memory. The first person to hit a brand-new subject sees the lite gauge quickly and the full card on their next visit.
4.3 How a verdict is made
Naming the subject. AI names the subject of the main query. Each result names the thing its page represents: rules first (a shop's product page → the product; IMDb → the film; an app store → the app; GitHub → the tool; a YouTube address → that video; a site's front page → the site), then one batched AI call for the rest (a review names its product; a news story names the article, found by link; a person is named as publicly known; a how-to page or a page about nothing names nothing, though a how-to typed as the query is a topic and gets a reading). Tracking variants of an address share a reading. Without an AI key an unnamed page falls back to a reading of the page, then of its website.
General versus recent. The summary and the bar's sentence describe how people regard the subject over time and never hinge on an incident the reader has no context for. When a development of the past two weeks — an outage, a redesign, a controversy, a price change — has clearly swung how people feel, with at least three dated entries behind it, one "Lately" sentence names it with roughly when it happened; otherwise nothing, and when in doubt nothing. A bare report that something is down right now is not an opinion of it.
A fair sample. The model reads a bounded sample taken in turns from each platform's most-reacted-to opinions, with no single thread allowed more than two fifths of the places, so a platform with thousands of entries cannot crowd out one with dozens. Search windows widen platform by platform until each has about 25 opinions or runs out of history.
Finding the discussion. By name on Reddit, YouTube, Hacker News and Bluesky; by link where the subject is an article or a video (Reddit's url: search, Hacker News's URL search, Bluesky's URL filter). Search windows widen from 3 to 12 to 36 months until enough is found. A bounded sample: at most a fixed number of threads per platform and comments per thread.
Two tiers.
| Tier | When | What it reads | Model | Cost |
|---|
| Lite gauge | for the bar, on first sight | counts, reaction scores, the top comments | a small, cheap model | a fraction of a cent |
| Full card | only on click | the full bounded sample | a capable model | a few cents |
Counting. The model classifies each opinion (positive / neutral / negative, on-topic or not). The split is counted from those classifications, weighted by reactions on a log scale. Fewer than eight on-topic opinions → "not enough to say", drawn as a grey bar or nothing. A confidence level with a one-line reason travels with every answer.
Articles (from v3). For a news story, "negative" means nothing on a story about a flood. The classification becomes stance toward the article — agrees with it, disputes it, reacting to the event, off-topic — and the bar is labelled "how people are reacting".
Memory. Each answer is kept for 24 hours, keyed by subject; served stale while a refresh runs in the background. Only derived data is stored (section 2, principle 4).
5. Components
| Component | What it is | Why it is needed | Cost |
|---|
| GitHub repository | This one: server/ and extension/ | Where the code lives; pushing deploys the server | Free |
| Server on Vercel | A small Next.js project: the gauge door, the card door, the embed page, a config door, a scheduled job | Holds the keys, does the thinking, remembers answers | Free tier; Pro (~$20/mo) at scale |
| Memory | Upstash Redis (via Vercel Marketplace) or Vercel KV | Answer once, serve everyone; keeps cost and platform usage down | Free tier |
| AI key | OpenAI API key; a small model for lite gauges and subject naming, a capable one for cards | Names subjects, classifies opinions, writes the sentence | Pay per use; cents per new subject |
| Source access | YouTube Data API key (exists); Hacker News and Bluesky need none; a new Reddit application (section 10); X optional and paid | Where the opinions come from | Free for v1 |
| The extension | Label, hands (Google skill), brain, drawer, icons; plain TypeScript bundled by esbuild with a small build script that writes a Chrome and a Firefox package from one source | The thing readers install | Free |
| Config door and kill switch | A small settings file the extension fetches on start: enabled/disabled per surface, current Google page selectors, request limits | Turn things off or patch selectors without a store update when Google changes its page | Free |
| Per-install token and limits | A random token made on first run, stored in the extension, sent with every request; the server limits per token, per address and per day | Stops someone scripting the doors to burn the AI budget | Free |
| Spend caps | Hard monthly limits in the OpenAI dashboard (and X's, if used) | A surprise becomes an outage, not a bill | Free |
| Golden set | Fifty subjects with hand-checked verdict ranges in server/golden/ | Every change to the analysis is run against them before it ships |
Out-of-pocket for v1: $5 plus AI usage (a few dollars a month at friend scale).
6. Sources
| Platform | Search by name | Search by link | Access | Notes |
|---|
| Reddit | yes | yes (url: in search) | off — no application has succeeded yet | Would be the best source for products and news. The connector is written and dormant; it returns when an application is approved. |
| YouTube | yes (costly) | a video's own comments by ID | existing key | A search costs 100 quota units of a 10,000/day free quota; reading a known video's comments costs 1. Request a quota increase before launch and prefer by-ID reads. |
| Hacker News | yes | yes | free, no key | Excellent for tech and business. |
| Bluesky | yes | yes | free, no key | Growing news audience. |
| X | yes | yes (url: operator) | paid, about half a cent per post read; the owner's token is connected | Read only for the full card (on click or the prefetch of the query's card), at most 20 posts per subject, with a default budget of 1,000 posts a day (X_DAILY_POST_BUDGET). |
| Facebook | no | a total share count only | app token, tightly limited | Counts, never posts. Not a source. |
| Instagram, TikTok, Threads | no | no | — | No public search door. Not sources. |
Re-uploads of a video (the file re-posted rather than linked) cannot be found through any door; that needs audio/visual fingerprinting of downloaded candidates, which the platforms do not permit at scale. A title-text search catches the lazy ones.
7. Version 1 scope
In
- Google results pages on the Google country domains (google.com, google.com.au, google.co.uk, …; Chrome match patterns cannot wildcard the country ending, so the label lists them).
- A bar beside results resolved to a named thing or a discussed article; nothing beside anything else.
- One bar for the search query itself above the results when the query names a thing ("sony xm6 review" → Sony WH-1000XM6); nothing for how-to or generic queries.
- The old site's look everywhere: the turquoise, shell-pink, peach and salmon palette and the glass surfaces on the bars as well as in the drawer.
- Hover sentence; click → the drawer opens over part of the page with the full card: summary, platform buttons, the bar, the recurring opinions, and the evidence view with the original posts, as on the old site.
- Lite gauge and full card; memory; config door with kill switch; per-install tokens; spend caps; golden set.
- Sources: Hacker News, Bluesky, YouTube. Reddit switches on the day the new approval arrives. X off.
- Chrome, loaded unpacked for development, then an unlisted store listing for a small beta.
- Privacy page.
Out (for now)
- YouTube pages, article pages, any other site.
- Bars inside other sites' link lists.
- Firefox, Safari, phones.
- Accounts, payments, comparisons, opinion over time, per-community breakdowns.
- Per-user Reddit login.
8. Roadmap
| Version | Adds | The new skill |
|---|
| v1 | Google results | hands read Google's result list |
| v1.1 | unlisted beta, a dashboard of cost per day | — |
| v2 | YouTube watch pages | hands read the video ID; server gathers the video's own comments plus every Reddit/X/HN/Bluesky post that shared the link (where it spread, when it peaked) |
| v3 | news and article pages on any site | hands read the headline and canonical link from the page's own labels; server searches by link (this article) and by headline (this story); stance instead of sentiment |
| v4 | every page | hands read the page's structured labels (product, film, app) on any site; an optional "everywhere" permission the reader opts into |
| v5 | side-panel query mode | the drawer answers "what do people think of your query" on every search without touching Google's page — cannot break when Google changes |
| later | Firefox (same build), Safari and iPhone (Apple developer account), share-sheet fallback for phones, nightly precompute index, X vs Y comparisons, opinion over time, "who says what" by community, an embeddable badge, a daily "guess the crowd" game for growth | |
9. Privacy and permissions
The browser warning: automatic bars and the floating site card require page access, which browsers describe as “read and change data”. There is no draw-only permission. The extension does not read passwords or form entries; it adds its own controls. Removing that access would require click-to-enable behavior and would stop automatic following. Turning the site-card switch off stops drawing, not the browser permission itself; readers can restrict websites in browser settings. No additional permissions or transmitted data were added for v0.14.
- Permissions requested at install: the Google search domains, every site (for the card on every site, below; the browser says so at install), and local storage for the install token, the config and the one-time note's flag.
- The card on every site (v0.11 as "take the bar with you", on by default from v0.12, on every site from v0.17.0): a site script runs on every page but Google's search pages and draws the badge. It asks the brain for the site's reading: the one carried from Google when the reader came from there (nothing is sent), else one fetched by the site's address — the origin (scheme and domain), never the page's path or query — and the name the site declares for itself (
og:site_name), which the brain then files for the site's other tabs. Carried and fetched readings live in session storage (gone when the browser closes). The × hides the badge on that page until it is next loaded; the menu switch ("Show the card on every site") turns it off everywhere, and the script then draws and sends nothing.
- "Analyse this page's subject" (v0.17.0): only when the reader presses the badge's second button, and after a one-time note, the page's visible text (reviews and comments first, up to about twelve thousand words), its title, its address and any JSON-LD it carries go to the page door once. Menus, headers, footers, scripts and everything typed into forms are skipped. The server reads the text and forgets it; only the finished reading is held, for a quarter of an hour.
- What is sent to the server, in all: the titles and addresses of results on Google search pages and the query; every visited site's origin and declared name (unless the card is off); a page's text on the button's press. No cookies, no account details, no page paths.
- Firefox's declaration (v0.17.0): the Firefox label names these in Mozilla's categories — "search terms", "browsing activity" (the site addresses) and "website content" (result titles and, on the press, page text) — under
browser_specific_settings.gecko.data_collection_permissions, which Firefox Add-ons requires of every new add-on and Firefox 140+ shows at install. Anything more sent to the server must be added there in the same commit as here.
- What the server keeps: derived answers for 24 hours; a finished card, quotes included, and the bar's sample (the entries it was counted from) for a quarter of an hour, so the list and the card agree with the bar; never a page's text; request logs without raw queries beyond what abuse prevention needs, for a short fixed period. The privacy page states the period.
- What the server sends to the AI: comment text and titles for classification, and on the press the page's text; never usernames.
- The store's forms: the "privacy practices" disclosure must match the above word for word; the description must match the function; no platform names or logos in the listing name or icon.
10. Reddit application (new)
The earlier search site has its own pending application with its own limits; it does not cover this project and its limits do not bind it. A new, separate application is submitted for the extension, describing it precisely. Mention the earlier ticket number and say this is a different product, so it is not treated as a duplicate.
Ask for exactly what the extension needs:
- Doors: search (
/search, with the url: filter for "every post that shared this link" and plain queries for "posts about this subject") and comments (/comments/{id}, top-sorted, shallow, capped per thread).
- Paging, with ceilings: up to 100 posts per link, 20 threads read in depth.
- Rate: all calls from the server under one identity; results shared through the memory; usage kept under the free 100/minute by design; rate-limit headers respected with back-off.
- What is kept: derived counts, sentence and links for up to 24 hours; verbatim excerpts fetched live so deletions propagate immediately.
- AI, stated plainly: comment text is sent to a model to classify and summarise; usernames are not; nothing is used to train anything (the OpenAI API does not train on requests by default).
- Hygiene: attribution and links to source threads;
over_18 content dropped; read-only; no user-profile doors; a public privacy page.
- Identity: register the app at
reddit.com/prefs/apps first; user-agent web:opinion-extension:v<version> (by /u/<account>).
- Structure: shared server credentials for v1; per-reader Reddit login is the upgrade path if the shared budget ever binds.
The Zendesk form hides its "Details of inquiry" box behind a theme quirk; the text must be entered through the page's editor or the submission fails as blank.
11. Store listing
| Store | Cost | Review | Covers |
|---|
| Chrome Web Store | $5 once | usually 1–3 days; longer for broad permissions | Chrome, Edge, Brave, Opera, Arc, Vivaldi |
| Firefox Add-ons | free | minutes to days; source uploaded alongside the build | Firefox desktop and Android |
| Safari | $99/yr Apple Developer Program, a Mac | about a week | Safari on Mac and iPhone |
Chrome requires two-factor authentication on the account, a verified publisher email, an EU trader/non-trader declaration (non-trader while non-commercial), a privacy policy address and the privacy-practices form. Every update is re-reviewed, faster. Start unlisted: reviewed and installable by link, not searchable.
12. Costs
Friend scale (tens of users): $5 once; AI a few dollars a month; everything else on free tiers.
Ten thousand daily users (~80,000 result pages a day, heavy repetition of subjects; paid tiers; rough, check current prices):
| Component | Load | Rough cost/month |
|---|
| Vercel Pro | about one request a second; easy | $20–60 |
| Memory | easy | $5–15 |
| Reddit | ~5,000 fresh subjects/day × 3 requests fits the free tier because of the memory | $0 non-commercial; ~$100 on the commercial tier |
| YouTube | free once the quota increase is granted and by-ID reads are preferred | $0 |
| Hacker News, Bluesky | fine | $0 |
| OpenAI | the real cost: reading ~50 comments per fresh subject and writing cards on click; model choice and caching swing this 5× | $500–3,000 |
| Total | | ~$600–3,200 |
At that scale the question is not whether the design holds (it does, to roughly 100,000 daily users) but who pays. Charging money or showing ads moves Reddit to its commercial pricing and the store listing to a business identity, so the monetisation stance is decided before ~1,000 users, not after.
13. What breaks first
| Risk | Likelihood | Mitigation |
|---|
| Google changes its results page; every reader's bars vanish at once | high, recurring | selectors served from the config door; kill switch; fail closed (draw nothing rather than draw wrongly); fast store updates; v5 side-panel mode does not depend on Google's page |
| YouTube search quota exhausted | certain at scale | quota increase before launch; by-ID reads; memory |
| AI bill surprise | medium | hard spend caps; lite tier by default; precompute popular subjects; daily cost dashboard |
| Someone scripts the doors | medium | per-install tokens, per-address and per-day limits, spend caps |
| Reddit rate limit | low with memory | back-off on headers; cache; per-user login as the upgrade path |
| Brigaded or bot-heavy threads | medium on news | log-weighted reactions; minimum sample; showing per-platform disagreement rather than one blended number; account-age signals later |
| Store rejection | low | narrow permissions; disclosures matching the code; no platform names or logos in the listing |
| Privacy complaint | low | least-privilege permissions; nothing sent from non-Google pages; a plain privacy page |
14. Build order
- Server first, no extension. The gauge door, the memory, subject naming with the rule table and the batched AI call, the lite gauge from Hacker News, Bluesky and YouTube. Tested from a plain web page that pastes in ten fake Google results.
- The extension. Label, hands with the Google skill, brain, drawer showing the embed page. Loaded unpacked in Chrome.
- Safety rails. Config door and kill switch, per-install tokens and limits, spend caps, the golden set.
- Public-facing. Privacy page, icons, screenshots, the unlisted listing. Friends install by link.
- Reddit on the day approval arrives — a new source behind the same doors, nothing else changes.
- v2 YouTube, then v3 articles, per the roadmap.
Each step is a few working sessions.
15. Repository layout and technology
opinion-extension/
|-- README.md this specification
|-- AGENTS.md rules for anyone (or any AI) changing the code
|-- .env.example the keys the server needs, with no values
|-- server/ Next.js on Vercel: doors, embed page, memory, cron, golden set
'-- extension/ WXT project: label, hands, brain, drawer, icons
- TypeScript throughout.
- Server: Next.js route handlers on Vercel; Upstash Redis for memory; Vercel Cron for the nightly job.
- Extension: plain TypeScript bundled by esbuild (Manifest V3; a Chrome package and a Firefox package from one source; Safari wraps the Chrome package in Xcode); Shadow DOM for anything drawn on a page; the drawer is an in-page overlay framing
/embed.
- AI: OpenAI through the Responses API with Structured Outputs, a small model for lite gauges and naming, a capable model for cards.
- Salvaged from the earlier search site: the YouTube, Reddit and X readers, the analysis prompt, the counting, the card UI and its stylesheet. Starting from scratch on the product; not on the parts that already work.
- No Node on the owner's machines: GitHub Actions installs, builds and tests every push and attaches the extension packages to the run; Vercel builds and runs the server from the
server/ folder.
Doors
| Door | Method | Purpose |
|---|
/api/gauge | POST { query, results: [{ url, title, snippet? }] } | Names the subject behind each result and the query, answers with a quick gauge per subject: from memory, computed within a 20-second budget, or pending with the work kept alive. |
/api/gauge?keys=a,b | GET | The poll for pending subjects: memory only. |
/api/card?key= or POST { key, context: { query, results } } | GET / POST | Full card computed on click within a bounded budget. POST recovers missing subjects from the original query or single result. X is read here only. |
/api/config | GET | The settings the extension reads on start, including the kill switch. Override instantly by writing JSON to the memory key config:override. |
/api/page | POST { url, title, site?, description?, text, data? } | One reading of the page the reader is on, from its subject: the page named by the model (a topic, a search or a listing is nothing in particular), then the platforms swept and the page read for its own reviews at the same time, then one pass over both; a page review weighs three platform posts. The page's own overall rating is read by a small call of its own, so a slow read of the page's opinions cannot lose it; that read is tried once more, shorter, when it does not land; and when neither lands the door answers 502 rather than publish a reading the page had no part in. Sent only on the badge button's press; the text is never kept; the finished reading is held 15 minutes. Needs the AI key. |
/embed?key= | page | The card the drawer frames. |
/api/health | GET | AI/source configuration and memory reachability; no credentials. |
/dev | page | Pretend to be the extension and knock on the gauge door. |
Every door answers any origin; the install token (X-Install-Token) and the rate limits are the protection. Without an AI key the gauge and the card still work, from a word-count estimate marked simulated, so the whole flow can be tried before a single key exists.
That word-count fallback applies only to main-query readings. Website/link reputation stays unavailable until AI is connected, since counting positive words cannot establish which site a post evaluates.
16. Decisions log
| Date | Decision |
|---|
| 2026-09-16 | v0.16.5 reverses hidden loading bars and knowledge-panel placement: all recognised results show a loading sweep immediately; the query gauge stays above the results and retains an empty state. Expanded-card controls wrap and the opinion count has its own line. Cached config is used only for its declared lifetime. |
| 2026-09-16 | v0.14: recurring opinions start collapsed behind a 23 px disclosure button; drawers have a separate, synchronized layer above bars. Soft no-scale hover, tooltip follows dragging, site pill grows to avoid clipping short names after updates. Rounded icons reverse the native theme mapping (cream/dark, grey/light). Publish all four browser packages by default. Keep existing automatic page access; explain its unavoidable browser warning instead of claiming draw-only access. |
| 2026-09-15 | v0.13.2 uses the owner's cream smile icon across all extension packages, with native grey/dark variants for Safari and Firefox and cream as the compatibility fallback. Site-card saves are serialized so simultaneous Google results cannot overwrite one another. |
| 2026-09-15 | Safari is packaged from the current Chrome source (v0.13.1, based on 8d8c41a), with the same UI and site card. Declare Safari's website permissions explicitly and document temporary folder loading on recent Safari. The old v0.3 Safari branch is superseded. |
| 2026-09-14 | Main-query opinion and individual URL reputation are separate; use an explicit domain fallback only when page-specific evidence is thin. |
| 2026-09-14 | Include sponsored results, product tiles and AI Overview references. Bare result bars; no main-card summary subtext. |
| 2026-09-14 | Full content loads only on expansion. Drawer is transparent frosted glass, sized to its content with a viewport ceiling. |
| 2026-09-14 | Deeper HN/Bluesky searches use full history and API-sized pages, retaining and disclosing bounded partial results. |
| 2026-09-14 | Accept CHATGPT/YOUTUBE deployment variable names; support exact YouTube-video comment lookup and cache-independent drawer recovery. |
| 2026-09-14 | Treat this as a fresh project: a browser extension, not a search page. New repository. |
| 2026-09-14 | Version 1 is Google results only; other surfaces are added as skills of the hands. |
| 2026-09-14 | The extension stays dumb; keys, thinking and memory on a server rented from Vercel. |
| 2026-09-14 | A new, separate Reddit application for this project; the earlier one's limits do not apply here. |
17. Trying it
Safari correction (v0.13.1). The Safari package uses the latest v0.13 source from 8d8c41a, including the current pill, drawer, menu and draggable site card. It replaces the obsolete v0.3 package from e983a18. Recent Safari can add the folder temporarily without Xcode; see SAFARI.md. The Safari label explicitly declares the website access already used by “take the bar with you”.
Nothing here needs Node on your machine.
The server. The existing Vercel project is opinion_meter, rooted at server/, at https://opinionmeter.vercel.app. The extension now uses that address by default. Set keys from .env.example (or the CHATGPT/YOUTUBE aliases), then redeploy. /api/health shows which are configured. Connect Upstash for persistent memory. Open /dev to test the gauge door.
The extension. GitHub → Actions → the latest ci run → download opinion-meter-chrome → unzip → chrome://extensions → Developer mode → Load unpacked → the unzipped folder. Then search Google. Details in extension/README.md. Edge takes opinion-meter-edge the same way at edge://extensions; Firefox takes opinion-meter-firefox at about:debugging; Safari needs the Mac app: follow SAFARI.md (download the opinion-meter-safari-app artifact from a run with the safari-app job, or build it in Xcode).
When a run is red. Open the failed step in Actions and paste the red lines into the conversation; that is the feedback loop that replaces a local Node.
The kill switch. In Upstash's data browser, set the key config:override to {"enabled":false} and every copy of the extension goes quiet within an hour (the config's ttlMinutes); delete the key to resume. The same key patches the Google selectors when Google changes its page.
18. Open questions
- Final name and whether to buy a domain.
- Monetisation stance (to be settled before ~1,000 users).
- Whether X is worth its cost for the full card, and at what daily budget.
- Which capable model writes the card, chosen by running the golden set against two or three candidates.
- Whether Chrome's built-in on-device model is good enough for a free "no-server" lite mode later.
