Tag: the-enshittification-machine-platforms-that-rot-after-winning

Posted on by

Subscription tiers that fragment formerly straightforward service guarantees

A service guarantee sounds simple until it becomes a product matrix.

Does the service include support?

Yes.

How fast?

That depends what you bought.

GitHub’s support structure makes the fragmentation unusually visible because the company publishes the response promises side by side.

Support became another upgrade path

In 2018, GitHub announced an expanded Premium Support program for Business Cloud and Enterprise customers. GitHub described the offering as a response to teams that needed around-the-clock care beyond its standard support. The new Premium and Premium Plus plans included 24/7 web and phone support, priority ticket handling, and guaranteed initial response times of 30 minutes for urgent requests and four hours for high-priority requests. Premium Plus added a named technical support account manager and other services. See GitHub’s 2018 Premium Support announcement.

The modern structure is even clearer.

GitHub Enterprise includes support, but GitHub’s current comparison lists the included Enterprise support level as 24/5 with an initial response of under eight hours for urgent and high-priority cases. Premium and Premium Plus provide 24/7 coverage and formal service-level agreements. Premium guarantees a 30-minute initial response for urgent tickets and four hours for high-priority tickets. Premium Plus adds a named Customer Reliability Engineer, incident management, technical advisory time, and faster normal-priority handling. See GitHub’s current Premium Support comparison.

Same platform.

Different promises about what happens when it breaks.

A response guarantee is not a resolution guarantee

GitHub is careful about an important distinction: its SLA covers the initial response, not the time required to solve the problem.

That is reasonable. A broken repository permission can be fixed quickly. A deep infrastructure bug may require engineering work regardless of how much the customer pays.

But the tier still changes the customer’s position during an incident.

One organization gets shared support during defined hours. Another receives 24/7 priority handling. Another has named personnel who already know the account and can coordinate an incident.

The service itself may have the same outage. The path through the outage is different because support has been packaged separately.

The guarantee becomes part of the upsell

There is nothing inherently dishonest about charging more for expensive human support. Dedicated engineers and 24-hour staffing cost money.

The Dead Internet Theory angle appears when a formerly straightforward expectation—I pay for the business service, therefore serious problems get serious support—splits into increasingly specific commercial tiers.

The customer is no longer buying only software capacity or features. The customer is buying a place in the support queue, an escalation path, and a promised response clock.

That changes how the cheapest paid tier should be evaluated.

The important question is not whether the plan technically includes support.

It is what support still means at that tier once faster response, stronger escalation, incident coordination, and dedicated expertise have all become separate things that can be sold above it.

Posted on by

Dispute systems shaped by the platform’s need to minimize handling costs

A $40 marketplace dispute cannot receive $4,000 worth of investigation.

That economic fact shapes the system before either buyer or seller clicks Open case.

Large platforms need rules that can process enormous numbers of disagreements without interviewing witnesses, hiring experts, or reconstructing every transaction from scratch. The result is usually a machine built around deadlines, predefined evidence, tracking records, and a small number of outcomes.

That is efficient.

It can also be brutally literal.

eBay reduces a messy transaction to evidence fields

eBay’s Money Back Guarantee is a good example. If a buyer says an item never arrived, eBay looks for specific evidence of delivery: an integrated carrier tracking number, a delivery or attempted-delivery status, date information, location matching the order details, and signature confirmation for sufficiently expensive orders.

If tracking shows delivery, eBay may close the case automatically. If a seller does not respond within the required period, eBay may step in. Buyers and sellers also have specific windows for asking eBay to review a dispute. See eBay’s Money Back Guarantee policy.

Payment disputes are similarly structured. A seller generally has five calendar days to accept or challenge a dispute and must submit the kinds of evidence the process recognizes. See eBay’s payment dispute guidance.

That turns a complicated story into something the platform can process at scale.

What gets lost at the edges

Standardization works best when the real-world event looks like the model.

Package delivered to the correct address with valid tracking? Easy.

Package scanned as delivered but left at the wrong building? Harder.

Buyer returns a different item, seller has incomplete photographs, carrier records are ambiguous, or an unusual transaction does not fit the expected evidence pattern? Now the dispute depends on whether the platform’s accepted fields can represent what actually happened.

An appeal exists, but it is also bounded. eBay says a buyer or seller can appeal within 30 calendar days and may need to provide new, additional information. eBay normally aims to respond within 48 hours. See eBay’s seller appeal process.

This does not prove eBay designed its rules with the stated motive of cutting support expense. eBay does not describe the policy that way.

The cost logic is an inference from scale: a marketplace processing millions of transactions needs repeatable rules, automation, and short review paths or the dispute system itself becomes uneconomical.

Cheap handling has a price too

Every rule that makes a case easier to process also decides what evidence counts.

That is where valid claims can disappear.

A system optimized entirely for individualized fairness would be too slow and expensive. A system optimized entirely for throughput would be a vending machine that occasionally eats somebody’s business.

The useful question is therefore not whether dispute resolution is automated.

It is whether the platform provides a meaningful escape hatch when the standardized model is wrong: a human review, permission to submit unusual evidence, and enough authority to reverse the automated result.

Efficiency is necessary at marketplace scale.

It should not become another word for the form has no box for what happened to you.

Posted on by

Marketplace sellers competing against the platform’s own offerings

Amazon can be the landlord, the checkout counter, the advertising system, the delivery network, and another seller on the shelf.

That combination does not prove the shelf is rigged.

It does create a conflict worth measuring.

Amazon opened its store to third-party sellers in 1999. The company says those independent sellers now account for more than 60% of sales in the store, and it argues that placing Amazon’s own offers beside third-party offers gives customers more selection and price competition. See Amazon’s response to the FTC’s 2023 antitrust lawsuit.

The same structure means Amazon is competing inside a marketplace whose rules, search systems, advertising products, fulfillment programs, and recommendation surfaces Amazon controls.

The conflict is about advantages, not mere coexistence

In 2023, the U.S. Federal Trade Commission and 17 state attorneys general sued Amazon and alleged that the company used its control of the marketplace to disadvantage competitors and sellers. Among the allegations, the FTC said Amazon had used recommendation widgets to promote its own private-label products and suppress competing information. The complaint also challenged seller fees and other marketplace practices. See the FTC’s case announcement.

Those are allegations in active litigation, not settled facts.

Amazon disputes them. The company says third-party businesses set their own prices, sellers can choose whether to use Fulfillment by Amazon or Amazon advertising, and Amazon’s own offers compete under systems designed to provide customers with low prices and fast delivery. Amazon argues the FTC misunderstands how retail competition works.

Both sides agree on the basic architecture: Amazon operates the marketplace and also sells products in it.

The disagreement is over what Amazon does with that position.

A platform has information ordinary sellers do not

The structural advantage is broader than a logo on a private-label package.

The marketplace operator sees search behavior, conversion rates, pricing, demand, advertising performance, fulfillment data, and the rules governing visibility. Individual merchants see only their own slice of that machine.

That does not mean Amazon necessarily uses every category of internal information to copy or suppress a seller. Claims like that require evidence for the specific practice.

But it does mean a seller is competing with an organization that also designs the playing field.

A normal retailer can choose which products receive shelf space. A normal marketplace can set rules for participants. Amazon does both at once.

That is why “self-preferencing” disputes matter even when customers like Amazon products and third-party sellers continue to grow. The issue is not whether the platform is allowed to sell things.

The issue is whether the operator’s control over discovery, fees, logistics, and marketplace rules gives its own offerings advantages that independent participants cannot realistically reproduce.

When the referee also has a team on the field, every important rule change deserves unusually close inspection.

Posted on by

Creators bearing more production risk under changing revenue shares

A streamer buys the microphone before the subscription revenue exists.

The same is true of the camera, computer, editing time, moderator help, graphics, music licenses, and the hours spent broadcasting to an audience that may or may not show up.

That is production risk. The creator pays much of it first.

On a platform such as Twitch, the reward side of that calculation also depends on rules the creator does not control.

Twitch changed premium subscription terms

In September 2022, Twitch publicly explained that its normal subscription revenue split was 50/50 on net subscription revenue, while some larger streamers had older premium agreements commonly described as 70/30 deals.

Twitch announced that affected streamers would keep the 70/30 split only on the first $100,000 of annual subscription revenue. Revenue above that level would revert to the standard split after the creator’s agreement renewed following June 1, 2023. Twitch explicitly acknowledged that some of those streamers had come to depend on the additional revenue. See Twitch’s 2022 letter on subscription revenue shares.

The creator’s production expenses did not automatically fall when that contract changed.

A studio built around an expected income level still has rent. Employees still need paying. Equipment already purchased does not become cheaper because a platform altered the percentage applied above a threshold.

That is how a revenue-share change shifts risk toward the creator: the production investment remains fixed while the rules governing its return can move.

Dependence makes the change harder to escape

A creator can leave Twitch.

That sentence is technically true and economically incomplete.

The channel’s followers, subscriber habits, emotes, moderation culture, discovery history, integrations, sponsorship expectations, and daily viewing routine may all be built around Twitch. Moving a video file is easy. Moving the social system around the file is not.

That dependence does not mean every change is exploitation. Twitch argued that the older premium agreements were inconsistent and disproportionately available to larger streamers. It also pointed to higher advertising revenue shares and other monetization systems as alternatives.

And the terms changed again.

In January 2024, Twitch removed the $100,000 cap on the 70/30 level for qualifying streamers and expanded its Plus Program with both 60/40 and 70/30 tiers. Twitch said the earlier cap had limited growth opportunities and acted as a disincentive. See Twitch’s 2024 payout-program update.

That later improvement is important because it proves the terms are not a one-way ratchet.

It also proves the larger point.

The creator owns the production bill. The platform owns the revenue architecture.

When a creator becomes heavily dependent on one platform, planning a business means estimating not only audience demand but also the chance that the platform may rewrite the economics after the audience has already been built.

Posted on by

Service reliability reduced while subscription prices increase

A subscription can get more expensive during the same period that the service has a bad reliability quarter.

That is irritating. It is not automatically proof that the company is charging more while permanently operating a worse system.

Microsoft 365 is useful here because Microsoft publishes enough numbers to separate those two claims.

The price went up

Microsoft announced a broad commercial Microsoft 365 price increase effective March 1, 2022. Business Basic moved from $5 to $6 per user, Business Premium from $20 to $22, Office 365 E1 from $8 to $10, Office 365 E3 from $20 to $23, Office 365 E5 from $35 to $38, and Microsoft 365 E3 from $32 to $36. Microsoft said the increase reflected the additional products and capabilities added to the suite over the previous decade. See Microsoft’s 2021 pricing announcement.

Microsoft raised commercial pricing again on July 1, 2026 for a range of Microsoft 365, Office 365, Business, Frontline, Windows, and related products. The company tied that increase to new security, management, storage, and AI capabilities. See Microsoft’s 2026 commercial pricing FAQ.

So the direction on price is documented.

Reliability has its own record

Microsoft also publishes quarterly worldwide uptime for Microsoft 365 business and enterprise services.

For 2022 and 2023, the published quarterly figures stayed between 99.98% and 99.99%. In 2024 and 2025, the numbers remained mostly in that range, although some quarters were lower, including 99.927% in the fourth quarter of 2024 and 99.954% in the fourth quarter of 2025.

Then the first quarter of 2026 dropped sharply to 99.526%.

That is a real documented reduction in availability compared with the surrounding quarters. It is not a screenshot from somebody whose Outlook stopped loading for ten minutes. It comes from Microsoft’s own service health and continuity data.

But the next number matters too.

Microsoft reports 99.994% worldwide uptime for the second quarter of 2026.

The service recovered.

A bad quarter is not a permanent decline

This is why “the service costs more and keeps getting worse” needs evidence.

The pricing trend is clear. The first quarter of 2026 was also clearly worse than Microsoft’s recent historical uptime figures. But the published data does not support saying Microsoft 365 entered a continuing reliability collapse. Q2 moved back above most recent quarters.

The more defensible complaint is narrower: paying customers can face higher renewal prices while still absorbing periods of materially worse availability, and the added features used to justify a price increase do not compensate a business at the moment mail, collaboration, or identity services are unavailable.

For a subscription platform, value is not only the feature list.

Reliability is part of the thing being purchased.

If a provider raises prices, the fair comparison is therefore not “more features than ten years ago.” It is what the customer pays now, what the service promises now, and how consistently the service actually works now.

Posted on by

Outsourced support with limited authority to solve actual problems

The support agent can understand the problem perfectly and still be unable to solve it.

That distinction gets lost whenever outsourced support is reduced to jokes about scripts and call centers. The interesting question is not whether the person on the other end is trying. It is what the system permits that person to do.

Airbnb provides a useful documented example because the company has publicly described how much of its support operation depends on outside providers.

In its 2022 annual report, Airbnb said the vast majority of its community support was performed by a limited number of third-party service providers. Those providers handled inquiries through telephone, email, social media, and chat. Airbnb also warned investors that the quality, staffing, training, and timeliness of third-party support could materially affect the host and guest experience. See Airbnb’s 2022 Form 10-K.

The contractor can handle the contact without owning the decision

A separate Airbnb outsourcing agreement filed with the SEC makes the structure more concrete. The service provider was contracted to operate contact-center services for guests, hosts, and other users. The statement of work included claims, resolutions, payments, trust, product specialists, safety, urgent support, and other support lines. See the Airbnb service-provider agreement.

That is not trivial work. Outsourced agents may be handling emotionally difficult cases, payment problems, reservation failures, and safety incidents.

But the same contract makes clear that the provider operates inside Airbnb-defined systems, access controls, procedures, staffing requirements, and service categories. Airbnb retains the ability to define the work, approve access, change services, bring work back in-house, or move it to another provider.

That creates a structural limit.

An agent may be able to gather evidence, explain policy, issue an allowed credit, or route a case. That does not mean the agent can rewrite the policy, override a trust decision, change a product limitation, or invent a remedy that the workflow does not authorize.

Escalation exists because first-line authority has edges

Airbnb’s current help system separately describes an internal complaint process for eligible business users in the EEA and UK. Those complaints are assigned to a case handler, and a business user who remains dissatisfied after the internal process can access mediation. See Airbnb’s business complaint process.

That separation matters.

A frontline support conversation and a formal review are not the same thing. One is optimized to handle enormous volume. The other exists because some disputes require somebody with broader authority to examine what happened.

Outsourcing is not automatically bad support. Specialized providers can supply language coverage, 24-hour staffing, and huge operational capacity that would be expensive to reproduce internally.

The failure happens when the platform makes the contact layer easy to reach but the decision layer nearly unreachable.

Then the user experiences a strange loop: every agent understands the problem, every agent is polite, every agent opens another case—and nobody in the conversation possesses the authority required to end it.

Posted on by

Declining moderation investment while audience size remains large

Moderation problems are easy to describe badly.

One abusive post does not prove a platform stopped moderating. One quick takedown does not prove the system is healthy either.

A better measure is capacity: how many people and technical resources are assigned to the job, how quickly reports are handled, and whether that capacity keeps pace with the size of the service.

X provides unusually concrete numbers.

The staffing cuts were measurable

In January 2024, Australia’s eSafety Commissioner published information supplied by X Corp. about staffing changes after the October 2022 acquisition of Twitter. According to the regulator, X reported a 30% reduction in global trust-and-safety staff, an 80% reduction in trust-and-safety engineers, a 52% reduction in directly employed content moderators, and a 78% reduction in global public-policy staff.

The same report said median response time to user reports about posts had slowed by 20%, while median response time for reports concerning direct messages had slowed by 75%. See the eSafety Commissioner’s summary of X’s transparency response.

Those figures are stronger evidence than screenshots of bad posts. They describe a documented change in the resources available to operate the system.

The audience did not become small

X was still describing itself as a massive service after the cuts. In its October 2023 company update, X said more than half a billion people visited the platform each month and reported 7.8 billion active minutes per day. See X’s own one-year post-acquisition update.

That does not mean half a billion people needed human moderation every month. Automation, policy changes, user reporting, and systems such as Community Notes can change how much work requires a person.

X has explicitly argued that its moderation strategy evolved rather than simply disappeared. The company has emphasized Community Notes, automated enforcement, and new safety work in particular areas. It later reported major enforcement activity against child sexual exploitation and said it was building a trust-and-safety center in Austin.

So the useful claim is not X stopped moderating. The documented record does not support something that simple.

The useful claim is that a platform serving hundreds of millions of people substantially reduced several categories of specialized safety staff and experienced slower report response times during the same period.

Moderation is part of service quality

For ordinary users, moderation is usually invisible until it fails.

Spam stays up longer. Harassment reports sit unresolved. Appeals take longer. Coordinated abuse becomes harder to investigate. The visible symptom may look like a content problem, while the underlying problem is capacity.

That makes moderation investment comparable to server capacity or customer support staffing. A giant platform can remain technically online while reducing the human systems that make participation tolerable.

The audience size is therefore only half the measurement.

The other half is how much infrastructure the platform still commits to governing the audience it already won.

Posted on by

Seller appeals systems that struggle to correct mistaken suspensions

A seller suspension is not the online equivalent of being asked to wait in the lobby.

For a business built around one marketplace, it can stop orders immediately, strand inventory, interrupt advertising, and turn every hour of review time into lost revenue. That makes the quality of the appeal process almost as important as the quality of the original enforcement decision.

Amazon’s current seller process shows how formal an appeal can become. Its guidance for a deactivated account directs the seller to Account Health, then Reactivate your account, where the appeal may require a questionnaire, an acknowledgment, a quiz, supporting documents, or additional information requested after the first submission. Amazon says sellers should normally expect a reply within two days. See Amazon’s account deactivation appeal instructions.

That is a real appeal path. It is also a path that begins after the account has already been deactivated.

Correcting the decision after the business stops

If an enforcement action is clearly correct, paperwork is mostly about showing that the underlying problem has been fixed.

A mistaken suspension is harder. The seller may need to prove that the premise of the action was wrong: the product was compliant, the tracking record was valid, the identity document belonged to the correct person, or a supposed policy relationship did not exist.

The appeal system therefore has to do more than collect another copy of the same information. Somebody or some process must be capable of reconsidering the original decision.

Amazon’s later Account Health Assurance program is interesting because it moves part of that work earlier. When Amazon announced the program in November 2022, it said eligible sellers whose accounts were at risk would receive proactive contact from an account-health specialist. As long as the seller could be reached within 72 hours and worked with Amazon to resolve the issue, the account would not be deactivated. Amazon described the program as a response to sellers wanting greater visibility into their account health and risk of deactivation. See Amazon’s Account Health Assurance announcement.

That changes the sequence:

possible violation → specialist contact → attempt to resolve → deactivation if necessary

instead of:

deactivation → seller loses access → appeal → wait for review

The protection is not universal

Account Health Assurance has eligibility requirements, and Amazon still reserves the right to act immediately on serious violations such as fraud or illegal activity. It is not evidence that every suspension outside the program is mistaken.

What it does demonstrate is that marketplaces can design enforcement around continuity while facts are being clarified, rather than treating shutdown first and review later as the only possible architecture.

A useful seller appeal system therefore needs more than a button labeled Appeal. It needs a clear allegation, relevant evidence requirements, a reviewer with authority to reverse the result, and enough speed that winning the appeal does not arrive after the business has already absorbed the punishment.

Posted on by

Customer support replaced by automated loops without effective escalation

Automation is excellent customer support until the customer’s problem is the thing the automation does not understand.

Account recovery makes the failure mode obvious.

Instagram’s official help system can walk a user through password resets, recovery addresses and linked-account options. Those paths handle enormous numbers of ordinary cases without requiring a human employee to read each request.

Then there are cases that fall outside the expected path.

The recovery tree can simply end

Instagram’s help documentation gives one unusually direct example. If a user has lost access to the email account used to register and did not link the Instagram account to Facebook, Meta says it is unable to give the user access to that account.

https://www.facebook.com/help/instagram/358911864194456

From the company’s side, that restriction has an obvious security purpose. Account recovery is an impersonation problem: a support agent who is too flexible can hand an account to the wrong person.

From the legitimate user’s side, however, the experience can be brutal.

The person may know the account’s history, photographs, old passwords, devices and contacts. The automated process still needs evidence in the forms the process was designed to accept.

When those forms are unavailable, repeating the same help flow does not add information.

It is just a loop.

Meta sells a route to a real person

The value of human escalation becomes clearer when the same platform explicitly offers it as a paid feature.

When Meta introduced Meta Verified in 2023, one of the listed subscription benefits was account support. Meta described it as “access to a real person for common account issues.”

Testing Meta Verified to Help Creators Establish Their Presence

Meta later expanded paid account-support benefits to businesses as well.

Expanding Meta Verified to Businesses

That does not mean every unpaid Facebook or Instagram user previously had a human agent and lost one. The support systems, account types and available channels have changed repeatedly, and some users can reach additional review processes.

The more precise observation is simpler: Meta itself recognizes that access to a person has value beyond automated help, because it sells that access as part of subscription products.

Good automation needs an exception path

Most support questions should not require humans.

People forget passwords. They misconfigure settings. They ask the same questions millions of other people have asked. Searchable documentation and automated diagnostics are faster for everyone when the case fits the template.

The design fails when there is no credible path for the case that does not fit.

An effective escalation route does not need to promise that the customer is right. It needs to let a qualified person inspect evidence the automated tree cannot evaluate, explain why a decision cannot be changed, or identify an actual system error.

Without that step, “support” can become a machine that repeatedly explains how to use the machine.

Very efficient.

Unless you need help.