why fakespot stopped working — and what fills the hole

On 1 July 2025, Mozilla shut Fakespot down. The official reason, in Mozilla's own 2025-05-22 announcement, was that the product "didn't fit a model we could sustain." Days later, the browser extension was gone and the website redirected to a goodbye page. This is the longer answer: what actually broke, why on-device AI is the part of the story Mozilla's blog post left out, and what fills the hole now.

TL;DR. Fakespot's fake-review detection ran server-side, which meant Mozilla paid for every review their users scanned. At consumer scale, that bill is roughly $10.50 per active user per month in inference costs alone. The unit economics don't close at a free price point — and a paid one shrinks the addressable market to the slice of shoppers willing to pay for a review-checker, which isn't most of them. So Mozilla pulled the plug.

Chrome 138 changed the math. As of Q1 2026, Chrome ships a browser-native Prompt API that gives extensions free access to Google's on-device Gemini Nano model. The AI now runs on the user's laptop, not the developer's server. Marginal cost: $0. Review Guardian is what Fakespot's replacement looks like when the inference economics finally work.

The shutdown — what happened, and when

Fakespot's ending wasn't a surprise to anyone watching closely, but it was a surprise to most of its users. Mozilla had acquired the company in 2023 with the explicit pitch that AI plus Mozilla's privacy posture could finally give online shoppers a trustworthy review-quality signal. Two years later, the model that did the detection — and the model in the financial sense — both stopped working.

The timeline, anchored to events that mattered for end users:

  • May 2016 — Fakespot launches as an independent product. Saoud Khalifah founds the company on the premise that machine learning can flag manipulated reviews on Amazon, Walmart, eBay, Best Buy and other large marketplaces. Free browser extension, free website grader, server-side scoring.
  • 2017–2022 — Fakespot grows steadily, attracting press coverage every time Amazon's review-manipulation problem makes the news. The user base climbs into the millions on Chrome and Firefox combined. Server-side scoring remains the architectural fact.
  • May 2023 — Mozilla acquires Fakespot. The announcement frames the deal as a values match: AI-driven review checking layered onto Mozilla's privacy stance, with Review Checker rolled out as an opt-in Firefox feature shortly after.
  • 22 May 2025 — Mozilla publishes "Investing in what moves the internet forward" on its corporate blog. Buried in the post is the announcement that Fakespot will shut down on 1 July 2025. The blog states bluntly: "While the idea resonated, it didn't fit a model we could sustain."
  • 1 July 2025 — Fakespot servers go dark. The browser extension stops scoring. The website redirects to a farewell notice. Review Checker is removed from Firefox in the same release cycle.
  • Late 2025 — A scattered set of replacement extensions appears on the Chrome Web Store — Savinoo, ReviewShield, SureVett, RateBud, FakeFind, Null Fake — most of them small solo-maintainer projects, each running its own server-side inference stack. None of them inherit Fakespot's user base. (install counts confirmed via Chrome Web Store listings)
  • Q1 2026 — Chrome 138 ships the Prompt API for extensions, exposing Google's on-device Gemini Nano model to any installed extension with the right manifest permissions. The model downloads once, runs locally on subsequent calls, and costs the developer nothing per request.
  • May 2026 — Review Guardian launches as the first MV3 Chrome extension to use Chrome's on-device Prompt API for Amazon review grading. The architecture Fakespot couldn't afford is now the architecture you ship by default.

None of this was sudden. The economics had been straining for years; the platform unlock arrived a few months after the product was put out of its misery. The gap between the shutdown and the unlock is roughly six months — short enough that a single well-positioned successor can still inherit the post-Fakespot search traffic before users settle on a long-tail of inferior replacements.

What Mozilla actually said

It is worth quoting the source directly because so much subsequent press coverage paraphrased it badly. The relevant passage from Mozilla's 22 May 2025 announcement reads:

We acquired Fakespot in 2023 to help people navigate unreliable product reviews using AI and privacy-first tech. While the idea resonated, it didn't fit a model we could sustain. — Mozilla, Investing in what moves the internet forward (22 May 2025)

Three things are worth pulling out of those two sentences.

First, "the idea resonated." Mozilla is explicitly saying user demand was not the problem. People liked Fakespot. People used Fakespot. The shutdown is not a "nobody wanted it" verdict; it is a "we couldn't afford to keep giving it to them" verdict. That distinction matters because it shapes what kind of replacement is viable. A replacement that fails for the same reason — server-side inference at consumer scale — will hit the same wall. A replacement that solves the economic constraint can pick up the demand that resonated.

Second, "didn't fit a model we could sustain." Mozilla doesn't say "model" lightly. The word is doing double duty: it gestures at the AI model (the thing that scored reviews) and at the business model (the thing that paid for the AI model to run). The two collapsed into each other. The cost of running the AI model was the business model problem. Larger AI vendors with deeper pockets — OpenAI, Anthropic, Google — might absorb that bill as a strategic loss-leader. Mozilla, a 1,000-person foundation funded primarily by a search-engine deal that has been under regulatory pressure, can't.

Third, "privacy-first tech." The acquisition pitch foregrounded privacy. But the shutdown reason is economic, not privacy-related. Read carefully, the post is not saying privacy was incompatible with the product; it is saying the price of doing the privacy-respecting version of the product was too high. That is a much more interesting failure mode, because it has a fix. The fix is moving the inference out of the data centre and onto the user's device — which was not technically feasible in 2023 when Mozilla acquired Fakespot, and is technically feasible in 2026 because of Chrome 138.

The honest version of Mozilla's farewell sentence, if it had been written in 2026 instead of 2025, would have been: "We acquired Fakespot in 2023 to help people navigate unreliable product reviews using AI and privacy-first tech. The unit economics didn't work — but they would now." Mozilla didn't write that sentence because writing it would have been an implicit endorsement of a competitor strategy. The replacement category got to write it instead.

The economic reality — why server-side review AI doesn't pay

To understand the shape of the problem, work the numbers Fakespot would have been running. A typical Amazon product page carries 50–500 reviews. Scoring it well — separating recycled language from genuine experience, weighting verified-purchase signals, modelling reviewer histories — needs more than keyword matching; it needs a real language model with enough context to read sentences in their setting.

How much does inference actually cost?

Use 2025 commercial pricing as a baseline. A general-purpose large language model from a major provider charges roughly $3.00 per million input tokens and $15.00 per million output tokens at the mid-tier rate, with cheaper "lite" tiers in the $0.50–$2.00 range when you are willing to accept lower accuracy. Reviews are short — call it 150 tokens per review on average — but a useful scoring pass also needs the prompt scaffolding (instructions, examples, schema enforcement) which adds another 1,000–2,000 tokens of input per call.

Score a single Amazon product page with 100 reviews:

100 reviews × 150 tokens         = 15,000 input tokens (reviews)
                                 +  2,000 input tokens (prompt scaffolding)
                                 = 17,000 input tokens

Per page cost (mid-tier model):
  17,000 input  × $3.00/1M       = $0.051
  ~500 output  × $15.00/1M       = $0.008
  -----------------------------------
                                  ≈ $0.06 per product-page scoring

Six cents per page. Sounds trivial. The catch is multiplication.

Multiplied across an active user base

An engaged Fakespot-style user does not score one product page per month. They score every product they are seriously considering buying, plus most of the ones they were curious about, plus the ones a friend linked them. Realistic numbers from the industry: 5 to 10 product-page scorings per user per active day. Active days per month: 15 to 20. Combine the medians:

Active user, monthly:
  ~7 product-pages/day × ~17 active days     = ~119 scorings/month
  119 × $0.06                                ≈ $7.14 per active user / month

…before you add:
  + retries on parse failures               (+10–15%)
  + re-scoring as reviews accumulate        (+10–20%)
  + sponsored-content / shipping-quality
    sub-prompts                             (+10–20%)
  -------------------------------------------
                                            ≈ $9–$11 / user / month

Call it $10.50 per active user per month as a round midpoint. That is the unit-cost wall Fakespot was hitting before payroll, hosting, support, fraud, billing, or compliance — pure inference. At 10 million users (a figure widely cited in industry coverage, though never published by Mozilla and therefore left here as a directional rather than primary-sourced number), monthly inference alone would have been $100M+. Even at one-tenth that user base, the bill is $10M a month.

Mozilla's entire annual operating budget across all programmes is in the hundreds of millions of dollars. The Fakespot inference bill alone — let alone a viable price-elastic business growing it — would have consumed a non-trivial fraction of that. It is not surprising that the strategic review concluded the line item couldn't be sustained. Any honest spreadsheet showed the same.

Why "just charge users" doesn't fix it

The obvious counter is to switch to a paid model. Charge $9.99 per month for the deep-AI tier, recover the inference cost, keep going. Three things break this in practice.

First, the conversion ceiling. Browser-extension freemium conversions cluster in the 1–5% range. Cookie Guardian, a privacy extension serving a similarly engaged audience, runs at roughly 11.5% — and that is considered an outlier. Even at an optimistic 5%, you need 20 free users to fund one paying user's inference. The free users' inference costs are still real. The arithmetic only closes if the paid tier price is high enough to cover the free tier's bill — which pushes the price out of the impulse-purchase band that browser-extension freemium relies on.

Second, the price-anchor problem. Fakespot had been free for the better part of a decade. Asking the legacy user base to start paying for a service they had associated with $0 is a brutal conversion. The natural reaction is to switch to a free competitor — and switching costs are zero, because it is a browser extension. Anyone who can solve the unit economics differently can offer the same product at the original price.

Third, the strategic-fit problem. Mozilla is, fundamentally, a browser company. Running an unprofitable AI inference operation as a side business, at a moment when the company is restructuring around Firefox and AI partnerships, is exactly the kind of line item a strategic review kills. The blog post is candid about this: "didn't fit a model we could sustain." The word "fit" is doing as much work as the word "sustain". Even if Mozilla could have made the maths close eventually, it wasn't the company they wanted to be running.

The pattern is general. Server-side AI for consumer free-tier products has a structural unit-economics ceiling, and it bites the moment usage scales past curiosity. Mozilla hit the ceiling at the moment they had to choose between subsidising it at scale and shutting it down. They chose the second option, which is the rational choice if no alternative architecture exists. By 2025, no alternative architecture existed at production quality.

The Chrome 138 unlock — what changed in early 2026

Chrome 138 shipped a feature called the Prompt API, exposed first to extensions and progressively to web pages. The Prompt API gives any extension with the right manifest permissions the ability to send a natural-language prompt to a small language model — Google's Gemini Nano — running entirely on the user's device. The model is downloaded once, cached on disk, and runs locally from then on. As Chrome's own documentation explains, "The network requirement is only for the initial download of the model. Subsequent use of the model does not require a network connection."

The shift is architecturally simple and economically enormous. The same review-scoring workload that cost Mozilla $0.06 per product-page on a hosted model now costs $0 per product-page for the developer of the extension — because the developer is no longer paying for the compute. The user's laptop is doing the work. Marginal inference cost: zero. The wall Fakespot ran into stops existing.

What the developer actually pays for instead

On-device AI is not free in the absolute sense. Three costs shift to different parties:

  • The user pays in compute. The Gemini Nano model is roughly a few hundred megabytes on disk, and inference uses the user's CPU/GPU for the duration of a call. On modern hardware (a 2022-or-later laptop) the latency is sub-second per review summary; on older hardware it can be several seconds. This is the actual tradeoff Review Guardian asks for: "your laptop does the AI work — that's the trade for $0 server costs and zero data leaving your browser." Most users accept the trade because they were already paying with their fans during a Zoom call.
  • Google pays in model development. Gemini Nano didn't appear from nowhere — Google built it, trained it, optimised it for on-device deployment, and shipped it inside Chrome. Google's incentive is platform leverage: extensions and web apps that use Chrome-native AI are extensions and web apps that don't run on Safari or Firefox. From a developer perspective, the upshot is that the model is available, the API is documented, and the price is zero. The model lock-in is the price Google charges; for a Chrome-extension developer that lock-in is acceptable because Chrome is already 65%+ of desktop browser share.
  • Chrome users pay in upgrade pressure. Chrome 138 is the floor. Users on Chrome 137 or earlier (or on a Chromium derivative that lags Chrome's release train) cannot run an extension that depends on the Prompt API. That is a real cost — but it is a cost that resolves itself, because Chrome auto-updates aggressively. By the time Review Guardian's listing is on the Chrome Web Store for a few months, Chrome 138 will be the floor for the vast majority of Chrome users, just as Chrome's auto-update history shows.

Why didn't Mozilla build this?

Mozilla could not have shipped the Chrome 138 unlock for a simple reason: Mozilla doesn't ship Chrome. The Prompt API is a Chrome-specific implementation tied to Google's Gemini Nano weights, distributed through Chrome's update channels. Firefox's WebExtension API does not have an equivalent endpoint at the time of writing, and Mozilla's strategic focus in 2025–2026 is on its own AI partnerships and Firefox-native features — not on building a competing on-device-LLM stack that would duplicate Google's investment.

The honest summary is: Mozilla was holding the right idea at the wrong technical moment, on the wrong browser, and the moment passed before the moment arrived. The category had to be picked up by someone building on Chrome, after Chrome shipped the platform primitive Fakespot would have needed in 2023.

The wider pattern. "Economic blocker just lifted" is a recurring shape for category resets. The cost curve for some load-bearing input falls below the price point at which the product can be free, and a category that died because of the previous cost curve becomes viable again — but only for a builder who shows up after the new curve is in place. Server-side AI inference at consumer scale was a wall. On-device AI inference at consumer scale is a doorway. Fakespot died at the wall. Review Guardian is built through the doorway.

What fills the hole — the on-device AI replacement

If you were a Fakespot user, the practical question is: what do I install now? The short answer is that no single replacement inherited Fakespot's user base. The longer answer is that the post-Fakespot landscape splits into three rough categories.

1. Server-side AI extensions — same architecture, smaller team

The largest visible group of replacements — Savinoo, SureVett, ReviewShield, RateBud, FakeFind, Null Fake — all run server-side AI inference on their own infrastructure. Architecturally they are smaller-budget Fakespots. Same model shape, fewer users, less press coverage. They are subject to the same unit-economics ceiling that ended Fakespot; the question is when their inference bill exceeds their willingness to subsidise it. Combined Chrome Web Store install counts across this category are well under 2,000 users at the time of writing — orders of magnitude below Fakespot's pre-shutdown scale.

From a shopper's perspective these are functional and most are free, at least for now. The privacy tradeoff is real: review text leaves your browser and is processed on their servers. If you're comfortable with that and the extension is still online when you read this, they work.

2. Browser-native built-ins — narrower scope

Chrome's own Shopping experience and Microsoft Edge's "Review Insights" both expose lightweight review-quality summaries on product pages. They are not architecturally similar to Fakespot — there is no separate trust grade, no per-review flagging, no whitelist UX, no manual deep-analyse trigger. They are integrated browser features designed for casual shoppers, not power-users. For the Fakespot workflow specifically (open the product, glance at the grade, decide whether to keep reading) they are not a substitute.

3. On-device AI extensions — the new architecture

Review Guardian is the first MV3 Chrome extension purpose-built on the Chrome 138 Prompt API for Amazon review checking. The architecture is exactly the one Fakespot couldn't afford and Mozilla couldn't ship: AI runs on the user's laptop, no review text leaves the browser, no server processes any data on the developer's behalf. Marginal cost: $0. Privacy posture: zero data collection, zero analytics, zero remote servers.

The detection runs in two passes. The fast pass is pure heuristics — verified-purchase ratio, time-clustering, reviewer history, rating distribution skew — producing an A–F trust grade on every product page in milliseconds, with no model inference required. The deep-analyse pass invokes Gemini Nano on a sample of reviews, looking for the language patterns generated text tends to exhibit, and adjusts the heuristic grade up or down. Both passes are free and unlimited for everyone — with no server bill to cover, there's no cap to sell.

This is the architecture Fakespot would have shipped if it had been founded in 2026 instead of 2016. The economic constraint Mozilla wrote about — the inability to sustain the model — is the constraint Review Guardian was designed around from day one.

Add to Chrome — Free See the full comparison →

Frequently asked questions

Did Mozilla say why they shut Fakespot down?

Yes — in the 22 May 2025 corporate blog post announcing the shutdown, Mozilla wrote that Fakespot "didn't fit a model we could sustain." They did not publish detailed financials, but the language strongly implies the cost of running the AI inference exceeded what the product could pay for. The shutdown took effect on 1 July 2025.

Is Fakespot coming back?

No. Mozilla's announcement was unambiguous about discontinuing the product, the website redirects to a farewell notice, and the Firefox Review Checker feature was removed in the same release cycle. The Fakespot domain and brand are owned by Mozilla and there has been no public indication of a relaunch. The realistic assumption is that Fakespot is permanently shut down.

Why didn't a server-side AI competitor just take Fakespot's place?

Several have tried — Savinoo, SureVett, ReviewShield, RateBud, FakeFind, Null Fake — but they hit the same unit-economics ceiling Mozilla hit. Running language-model inference on millions of reviews per day costs real money per user, and the freemium price points that work in browser-extension distribution don't cover the bill. The combined install counts of these replacements remain well under 2,000 users — they are absorbing scattered demand, not inheriting Fakespot's market.

What's actually different about an on-device AI extension?

The AI model runs on your laptop, not on a server somewhere. That has two consequences. First, your review text never leaves your browser — there is no server log of which products you scanned, no analytics, no remote data collection. Second, the developer's marginal cost per scoring is $0, which means the product can stay free at scale without the unit-economics problem that ended Fakespot. The tradeoff is that your laptop does the compute, and the extension only runs on Chrome 138 or newer.

What's the catch with Chrome's Prompt API?

The model is downloaded once, then runs locally. The download is a few hundred megabytes, which is meaningful on a metered connection but a one-time event. The model is Google's Gemini Nano, which means the extension is Chrome-only (it won't run on Firefox or Safari today). The latency depends on the user's hardware — sub-second on a modern laptop, a few seconds on older devices. None of these constraints are deal-breakers for the shopping use case, but they are the honest tradeoffs.

Is Review Guardian a Fakespot replacement?

Yes, in the literal sense: it grades Amazon product reviews on a letter scale, flags suspicious patterns, and runs as a browser extension that activates on Amazon product pages. The architectural difference is that it uses Chrome's on-device Prompt API rather than server-side inference, which is what makes the free price point sustainable and the privacy posture genuine. The product is Amazon-only at launch (one DOM to maintain well rather than five DOMs to maintain badly); Walmart and other marketplaces are on the roadmap as the on-device model and the parser mature.

What happens to Fakespot's old browser extension installations?

The extension still appears in the Chrome Web Store and in users' extension lists, but the backend is gone — the icon either greys out or returns an error when invoked, depending on which version is installed. Mozilla has not committed to removing the listing. Practically, the right move is to uninstall it, both to free the toolbar slot and to remove a permission grant for an extension that no longer functions.

What this means for review-checking in general

Zoom out. Fakespot's ending is not a story about one company; it is a story about a category whose economics were broken from the outset and which had to wait for the platform underneath it to catch up. Three lessons fall out of the transition that are worth pricing into whatever you replace Fakespot with.

  • Architecture is the moat. The replacement extensions that ship with server-side inference are running the same economic experiment Mozilla ran. Some will fail the same way. The replacements that ship on Chrome's on-device Prompt API don't have the same exposure, because they don't have the same bill. When picking a successor, the architectural question — where does the AI inference actually run? — is more predictive of long-term survival than the feature list.
  • Privacy posture is no longer a tradeoff. The historical assumption was that AI quality required server-side compute, and server-side compute required sending the user's data to a third party. Chrome 138 broke that assumption. An extension that does AI scoring without exfiltrating review text is no longer making a feature compromise — it is using the platform the right way. Review Guardian's "no servers, no tracking, no account" stance is not marketing; it is what the architecture forces.
  • Browser-platform shifts are a category reset, not an evolution. Cookie AutoDelete died when Chrome enforced Manifest V3 and the volunteer maintainer didn't ship an MV3 port; the category reset, and the survivor was the extension built MV3-native from day one. Fakespot died when server-side AI inference economics stayed broken too long; the category reset, and the survivor will be the extension built on the Prompt API from day one. The pattern repeats. Pick the extension that started after the reset, not the one that survived through it.

None of this is an indictment of Fakespot, which was a well-built product that solved a real problem for a long time. It is a property of the model — the AI model and the business model both — that the work the product did got cheaper to do over the same years that it got more expensive to scale. Mozilla's strategic review caught the crossover. The replacement category is built on the cheaper side.

Sources and further reading