Websites replay your browsing: the business behind your clicks

On an online shop, your clicks, pointer movements and some form entries can be recorded to reconstruct your visit. A member of staff can then watch the journey unfold again. This technique is called session replay. Its capabilities vary with the software, its settings and the safeguards applied. ^L01

A visit ends at the sign-in page. In the retailer’s analytics, it becomes another exit from the checkout. With session replay, the team can return to the sequence of actions: the field selected, the button pressed, the message that appeared, then the departure. That second look explains the appeal of the product. It also helps explain how much detail it can collect.

On November 5, 2026, France’s data protection authority, the CNIL, published its final recommendation on these tools, adopted on October 8. It acknowledges their value in diagnosing errors, improving journeys and assisting customers. It also sets out the conditions for consent and calls for collection proportionate to a purpose defined in advance. The document explains the existing framework and suggests ways to comply; it expressly describes itself as non-regulatory guidance. ^{L01 L02 L03}

The economic question begins here. A better understanding of a problem may lead to a profitable fix. Between the recorded scene and the claimed benefit, however, lie a diagnosis, design work and a measurement of results. The data may also serve purposes of the supplier’s own. To assess the value of a recorded visit, we need to follow both the business decision and the personal information behind it.

Rebuilding a page, one event at a time

In the web architecture documented by rrweb, the browser first produces a description of the page. It then records changes and timestamped events. A session player applies them in sequence to reconstruct the visit. The moving page is generated from structured data, without capturing a video stream. Mobile software can use other methods; the description here concerns this particular web architecture. ^{T01 T03}

The structure being observed is the DOM, or Document Object Model: the organization of page elements as the browser works with them. It includes headings, forms, buttons and text replaced after an action. A collector can capture an initial state and subsequent changes: a message appearing, a value changing, a pointer moving or the page scrolling. The exact capabilities and exclusions depend on the product and its configuration. ^{T02 T03}

This gives a team context around an error. A counter may show how often something went wrong; a replay can help explain the circumstances. Perhaps a message was too far down the page, a window covered a control or an unexpected state blocked progress. These are possible uses, each requiring verification on the interface concerned. The CNIL recognizes this contribution of individual observation when aggregate metrics leave important context unexplained. ^L04

Method and scope

The diagram represents rrweb’s documented web architecture: an initial state plus timestamped events and changes reconstruct the page. Filtering is illustrated before transmission. Actual capabilities and exclusions depend on the product and its settings. Mobile implementations may differ.

Collection can take place while someone is typing. A script observing changes to a field may access its value before the form is submitted, subject to the protections in place. The “send” button triggers a business action. By itself, it seldom marks the boundary of everything that code can read on the page. The rrweb documentation and research on online forms establish this distinction. Understanding collection therefore requires examining the instrumentation and its settings, beyond the visible interface. ^{T02 T04 T11}

A replay also inherits the limitations of its construction. Missing resources, unusual components or omitted events can leave gaps. What appears in the player is a representation assembled from the retained elements. Its accuracy deserves checking, particularly when it is being used to explain a rare fault or support an important business decision. ^{T03 T07}

Masking and the fate of recorded data

Protection often starts in the browser, with masking before transmission. Fullstory explains that, when enabled, its Private by Default mode replaces text before transmission unless capture has been explicitly allowed. It also distinguishes excluding an area, which removes more elements and interactions, from masking its text while potentially retaining layout and gestures. Clarity likewise documents masking before transmission. These are suppliers’ descriptions of their mechanisms; the behavior of a particular website depends on how they are integrated. ^{T06 T08}

The CNIL uses “masking” in a broader framework. Its recommendation distinguishes three arrangements. Data may be collected with unmasking subject to authorized access and internal approval. It may be collected under encryption, with the supplier providing keys after a justified request. Or the data may remain outside collection altogether. These arrangements are labeled M1, M2 and M3. They sit alongside other measures recommended for the relevant purpose, including masking by default and blocking especially risky content. ^L02

The distinction changes the exposure. Keeping a value out of collection removes that value from subsequent access to the collected dataset. An encrypted value remains recoverable by whoever holds the means to decrypt it. A value protected through an unmasking procedure depends on access permissions and the scrutiny applied to requests. The choice affects security, day-to-day operations and the scope for later reuse.

Product names require the same care. Contentsquare’s documentation, updated on October 26, 2026, describes masking before data leaves the browser and, separately, an option to collect selected content under encryption. In that option, the customer’s private key enables decryption. The key arrangement differs from the CNIL’s M2 example. Comparing the two requires establishing who actually holds the key. ^T09

Information also needs to be followed across the page. A protected field may be copied into a summary, a heading, a page address or a custom property. This is a risk scenario whose realization depends on the site and the tool. Contentsquare explicitly distinguishes protections across channels and identifies values for which the integrator remains responsible. An interface change may therefore require a fresh check of the rules. Maintaining protection is part of maintaining the product. ^T09

Documented risks, dated evidence

In November 2017, Princeton researchers reported detecting recording by the services they studied on 482 of the top 50,000 Alexa-ranked websites. They described the finding as a lower bound associated with their method. It is a historical observation for that sample and those services. Measuring current prevalence in France would require a contemporary investigation. ^T10

Leaky Forms, published at USENIX Security in August 2022, drew principally on data collection in May and June 2021. The researchers identified 52 websites where session replay scripts inadvertently captured passwords. They disclosed the problems and reported remedial changes. A separate finding on email addresses transmitted before submission covered a wider range of tracking technologies. Combining the two would attribute observations to session replay that extend beyond its scope. ^{T11 T12}

More recent work examines mobile applications. A public version of TapSpy, accepted for CCS 2026, studies Android integrations through static analysis and language-model-assisted reasoning, with partial dynamic validation. The conference is scheduled for November 15–19, after this investigation’s cutoff. Risks classified by that method must be distinguished from directly observed leaks. We do not use them to estimate a leakage rate for the French web. ^{T13 T14}

Together, these studies document risk mechanisms and observations within defined boundaries. They support checking actual collection, exclusions and the durability of those exclusions as software changes. The material assembled here provides no basis for a monetary estimate of average harm per recorded visit.

A rapid click leaves room for interpretation

Analytics interfaces use the term rage clicks for rapid, repeated clicks within an area. The suggested interpretation is obvious: something has irritated the visitor. Yet Fullstory documents false positives, including calendar controls designed to receive repeated clicks. The same pattern can be part of ordinary use. ^{T15 T16}

A team has several links to establish. Does the action reveal a difficulty? Does that difficulty explain an exit? Was the visitor intending to buy? Would a change help them finish? Each question needs its own evidence. A persuasive recording may demonstrate an interface defect while leaving the scale of its commercial effect unresolved.

Selection adds another layer. Which people can be observed depends on consent, the pages instrumented, technical compatibility and any sampling applied. The analyst then chooses cases that attract attention. Our methodological inference is straightforward: a striking scene demonstrates a possibility; estimating its frequency across traffic requires a documented denominator.

Measurement can take the form of an A/B test: randomly assigning eligible visitors to the existing version or a revised one, using a metric defined before examining the results. Research on online experiments explains the importance of population selection, sample sizes and measurement rules. The associated collection protocol must also meet the applicable requirements. ^{R01 E14}

A sound test can estimate the effect of the change. Evaluating the instrument raises a further question: would another method have found the same defect, at the same time, and at what cost? Replay may speed up bug reproduction or help teams agree on a diagnosis. Attributing the whole benefit of a redesign to the tool would require examining that alternative. This distinction runs through our assessment of its economic value.

The boundaries around a conversion claim

Contentsquare’s Diptyque case describes such a sequence. Journey and click analysis led the team to investigate the checkout sign-in interface. Replays helped refine the diagnosis. Two variants were then tested with AB Tasty. The supplier reports a 9% increase in mobile conversion on the page, defining conversion as reaching the shipping-address entry step on the French and US sites. ^E01

That definition places the result at a particular stage. The page supplies neither starting rates nor sample sizes, an uncertainty interval or a precise testing calendar. It also leaves the effects on completed orders and contribution margin unreported. We therefore retain the figure as a supplier-reported commercial result, alongside its definition. More visitors reaching the address step could herald a useful gain; establishing the effect on purchases requires another measurement.

At MOO, a company blog post dated October 11, 2024 confirms the use of Fullstory alongside interviews and customer feedback. Fullstory’s commercial case study gives a drop-off rate falling from 41% to 29% in the optimized journey. Those levels imply a reduction of 12 percentage points, or approximately 29.3% in relative terms. The same document nevertheless highlights a “12%” decline. The calculation and clarification of units are ours. ^{E04 E05}

The customer’s own account confirms use of the tool. The numerical outcome remains a supplier-published claim, with no observation window, sample sizes or control group allowing the evaluation to be reproduced. These cases give the reader something concrete: a problem, work undertaken and a reported result. Their reach remains local, with uneven access to methods and underlying data.

Misreading the numbers creates an economic error. Applying an improvement in an intermediate step to total revenue turns a journey metric into assumed income. Subtracting percentages without examining their starting levels can alter the apparent size of a result. The denominator deserves as much scrutiny as the dashboard presenting it.

Clarity and the second use of visitor data

Microsoft presents Clarity as a free service and states in its FAQ that it does not sell data. It also describes itself as a data controller. Its technical documentation provides a more specific account of one use: an advertising consent signal governs sharing with Microsoft Ads. ^{T07 E09}

In Microsoft’s table, updated on May 21, 2026, ad_storage=granted enables that sharing to improve retargeting and campaigns. The denied state blocks sharing with Microsoft Ads. This documents a conditional mechanism. It establishes neither activation on a particular website nor revenue earned from an individual visit. Our analysis rests on this directly accessible evidence; it makes no assumption that recordings are sold. ^E09

This architecture connects with the economics of personal-data collection: the same information can have different values for different users. For a retailer, it may help resolve a problem or reduce time spent on support. A supplier with purposes of its own may use data in other products or services under the applicable conditions. A software invoice of zero therefore leaves questions about purposes, recipients and integration costs. The public documents reviewed offer no average monetary value for this second use.

Method and scope

Microsoft Learn describes limited cookieless collection when analytics_storage is denied, and Microsoft Ads sharing conditional on ad_storage. Another page describes cookieless replay as one page per session. These documents establish the stated behaviour, without establishing any particular site’s settings or legal compliance.

Another detail affects measurement itself. Microsoft explains that, in the mode described without cookie consent, each page view becomes a session and each recording shows one page. Continuity across pages is lost. Its more recent consent-management page describes cookieless collection with limited functionality when analytics_storage is denied. These documents set out the product’s stated behavior. We have not instrumented a website to examine a particular deployment. ^{E09 E10}

An analytical consequence follows: changing a setting can alter the session count and apparent progression rates. Before diagnosing a deterioration in business performance, a team needs to check that the measurement system remains consistent. The same journey may have been divided differently.

The phrase “cookieless” also requires looking at what actually happens on the device. Article 82 of France’s Data Protection Act covers accessing and storing information, and the CNIL requires prior consent for the replay purposes it examines. A technical mode’s name cannot settle its legal classification. Assessing a deployment would require its actual behavior, purposes and disclosures to visitors. ^{L02 L06}

How much improvement would pay for the work?

For a retailer, the decisive calculation compares additional orders with the costs incurred to obtain them. Contribution margin per order means here the amount remaining from an extra order after the relevant variable costs: the product, payment processing, fulfillment, subsidized delivery or returns, depending on the business. That definition must stay consistent between the groups compared.

Our teaching model simplifies the calculation as net incremental contribution = N × Δp × m − C. N is the relevant traffic exposed to the correction over the period; Δp is the absolute change in the completed-order rate; m is the average contribution per additional order; and C is the cost included for the work over that same period. This is a scenario identity whose usefulness depends on the inputs. N and the denominator of the order rate must use the same unit of observation. ^M01

Costs include any subscription, integration, time spent reviewing sessions, the fix itself and the maintenance of safeguards. A one-time cost needs an explicit allocation to the chosen period. The calculation must also identify the decision being made: a new installation, a diagnostic project or an extension of use. A cost already incurred and unchanged between alternatives should not be counted a second time in comparing them.

Consider an entirely hypothetical example to make the units clear. With 100,000 visits exposed to a correction, €30 in contribution per additional order and €3,000 in costs over the same period, breaking even requires 100 extra orders. The purchase rate must rise by 0.10 percentage points. From a starting rate of 2%, that means reaching 2.10%, a relative increase of 5%. These assumptions describe neither Diptyque nor MOO, and they are no estimate of an average tool’s return. ^M01

Method and scope

Net gain = visits × contribution per order × (absolute purchase-rate increase in percentage points / 100) − cost. Here 100,000 visits, €30 contribution and €3,000 cost imply 100 extra orders and 0.10 percentage point at break-even. A 2% initial rate reaches 2.10%, a 5% relative rise. The period and cost scope must remain consistent; assumptions require compatible contribution and an effect applicable to exposed traffic.

N refers to traffic receiving the change. It may differ from the number of sessions recorded to discover the problem. Applying an estimated effect to a different population, however, requires justification. Changes in basket size, returns or service costs can also make the model’s average margin inappropriate. In that situation, directly comparing contribution per randomly assigned visitor provides a more relevant measure.

Uncertainty around Δp remains crucial. A positive result based on few orders may be compatible with a very small, or even adverse, effect. The financial threshold helps frame the question; the experiment must then estimate the effect and its uncertainty. Research on online testing documents pitfalls in the interpretation of metrics and populations. ^E14

Consent becomes part of the product

The French recommendation distinguishes two stages. Reading and writing information on a device fall under Article 82. For the replay purposes examined, the CNIL concludes that prior consent is necessary. Processing the personal data obtained must then have a lawful basis under the GDPR. According to the authority, consent will generally be the most appropriate basis, subject to context. Establishing a basis for that second stage leaves the first requirement in place. ^{L02 L06 L07}

The reasoning starts with the service the person requested. Buying a product requires certain information to fulfill the order. A detailed recording of the journey gives the retailer an additional diagnostic benefit. In its response to consultation submissions, the CNIL distinguishes that benefit from what is strictly necessary to provide the visitor’s service. ^L04

This makes the disclosure important. A general reference to audience measurement may suggest an aggregate counter to a reader. The CNIL recommends specific information about replay at the first layer of the consent interface, separate from the purposes presented. Users should be able to understand that an individual sequence of navigation will be recorded and viewed. Withdrawal must also bring the relevant reading and writing operations to an effective stop. Collected data must be deleted where no other lawful basis justifies retaining it; the processor must receive the corresponding instruction. ^{L01 L02 L05}

A supplier can occupy different roles. When it processes data solely for purposes defined by the website publisher, it acts as a processor. When it uses data for its own purposes, responsibilities change for those operations. The CNIL describes, among other situations, possible joint controllership for device access that serves both parties. This assessment follows the actual uses of data and affects the contract, evidence of consent and arrangements for exercising rights. ^{L02 L07}

A session associated with a pseudonym may still be personal data. Under the GDPR, identification is assessed in light of the means reasonably likely to be used. A detailed history, information that can be combined with other sources or an identifier linked to an account therefore needs examination in context. Demonstrating anonymization requires more than replacing a displayed name. ^L07

Retention follows the purpose

Retention and deletion of data copies depend on the systems holding them and the purpose still being served.

The periods a product offers and the periods a business needs answer different questions. Microsoft documents 30 days for Clarity playback data and nine months for labeled or favorited sessions. Those product settings describe availability. An organization’s appropriate retention period depends on its purpose and the associated risk. ^{E11 L07}

The CNIL’s annex gives purpose-specific benchmarks: a few months to understand errors or work on usability in the current website version; a few hours after the session for support, with documented justification taking account of when requests arrive. The recommendation allows room for alternative measures that meet the obligations. These benchmarks should be read alongside their purpose, rather than converted into a single retention period for every use. ^L02

Ordinary product changes also test the arrangement. A new form, a conversation with an assistant or an added summary may change the data accessible to a collector. Access permissions evolve as teams change. Deletion must work at the relevant level, and people’s requests must be handled using the available information without routinely creating a permanent identifier solely for that purpose. The GDPR specifically addresses processing that does not require identification. ^{L02 L07}

The diagrams in this investigation use synthetic examples. The contribution chart uses the fictional assumptions detailed below. These illustrations explain mechanisms; they remain passive and record no reader interactions.

Funding a better decision

The public evidence documents a useful role for replay: retaining enough context to understand an interaction, share a diagnosis and decide on a correction. Commercial cases show local outcomes within a broader process. They leave open the instrument’s average return and the share of the benefit that a less intrusive method might also have delivered.

Investment may make sense when a team has a specific problem, the capacity to act and a way to evaluate the outcome. Accumulating sessions without a defined question can increase the burden of reviewing, protecting and retaining them while leaving their expected value uncertain. That is an inference from our analysis of the mechanisms, rather than a statistical estimate of the market.

The CNIL’s recommendation brings the issue back to a practical economic requirement: define the information needed before collecting it. Value develops through a better decision. Measuring it requires knowing which data helped, which change was made and which final outcome actually changed.

Terms used in this article

DOM: the structure of page elements managed by the browser. Session replay: reconstruction of an individual journey from recorded elements. Masking: a protection whose mechanism must be specified; content may be replaced, retained under restricted access or excluded. A/B test: comparison of variants randomly assigned to a defined population. Percentage point: the difference between two rates expressed as percentages. Contribution margin per order: revenue remaining after the variable costs included for an additional sale. Pseudonymization: replacement or separation of identifiers while a possibility of linking data to a person remains. CMP: the interface and system used to manage consent choices.

Research method and calculations

Documentary research closed on November 8, 2026, Europe/Paris. We reviewed CNIL texts, technical documentation, original research and supplier-published cases. MOO’s use of Fullstory was checked against MOO’s own website. Commercial performance claims remain attributed to their publishers. Clarity pricing and terms pages available only through search indexes were excluded from the article’s contractual conclusions. No third-party site was instrumented, no participants were interviewed and no conversion experiment was conducted for this investigation.

M01 · l0g calculations. MOO difference: 29 − 41 = −12 percentage points; relative change: (29 − 41) / 41 × 100 = −29.268…%. Fictional scenario: break-even Δp = C / (N × m), with Δp expressed as a fraction; the result in percentage points is 100 × Δp. With N = 100,000, m = 30 and C = 3,000, Δp = 0.001, or 0.10 percentage points. At a starting rate of 0.02, the relative increase is 0.001 / 0.02 = 5%. The model assumes one completed order per converted visit, a common period, a compatible average margin and an effect applicable to the traffic considered.

The cover illustration is conceptual artwork generated with an AI tool. It depicts a reconstructed browsing journey and does not reproduce any supplier’s interface.

Sources and documents

1. L01 · CNIL · Accessed: 2026-11-08Outils de rejeu de session : la CNIL publie sa recommandation pour guider les acteurs
   Passages and scope

   Published 5 November 2026; adopted 8 October 2026.

   Announcement of the final recommendation and changes following consultation.

   Regulator summary; the full recommendation specifies the scope of its measures.

2. L02 · CNIL · Accessed: 2026-11-08Outils de rejeu de session ; Recommandation
   Passages and scope

   Adopted 8 October 2026; published 5 November 2026; 12 pages.

   Consent and legal bases (pp. 6–8), M1/M2/M3 masking arrangements (p. 10), security and retention (pp. 11–12).

   Non-regulatory guidance. Distinguish existing legal duties from recommended implementation; alternatives may be justified.

3. L03 · Journal officiel de la République française / Légifrance · Accessed: 2026-11-08Délibération n° 2026-114 du 8 octobre 2026 portant adoption d'une recommandation concernant les outils de rejeu de session
   Passages and scope

   Decision adopted 8 October 2026; Official Journal publication 5 November 2026.

   Official record of adoption and publication of the recommendation.

   Read alongside the CNIL annex; adoption and publication are different dates.

4. L04 · CNIL · Accessed: 2026-11-08Consultation publique sur le projet de recommandation concernant les outils de rejeu de session ; Synthèse des contributions et réponses de la CNIL
   Passages and scope

   Document dated October 2026, linked from the 5 November announcement.

   Consultation arguments and the CNIL’s responses, including the value of individual context.

   Voluntary contributions, rather than a representative sample of the market.

5. L05 · CNIL · Accessed: 2026-11-08Rejeu de session : ces outils qui enregistrent votre activité en ligne
   Passages and scope

   Published 5 November 2026.

   Visitor information, refusal and withdrawal of consent, and an overview of rights.

   Public guidance to be read alongside the legislation and recommendation.

6. L06 · Légifrance · Accessed: 2026-11-08Article 82 de la loi n° 78-17 du 6 janvier 1978 relative à l’informatique, aux fichiers et aux libertés
   Passages and scope

   Article 82, version in force since 1 June 2019, accessed 8 November 2026.

   French Data Protection Act: access to and storage of information on a terminal, information and consent.

   Application to a service depends on its actual operations and purposes.

7. L07 · Journal officiel de l'Union européenne / EUR-Lex · Accessed: 2026-11-08Règlement (UE) 2016/679 du Parlement européen et du Conseil du 27 avril 2016 (RGPD)
   Passages and scope

   Adopted 27 April 2016; published 4 May 2016; applicable since 25 May 2018.

   GDPR: purposes, minimization, legal bases, rights, roles and security; Articles 5–7, 11, 25–28 and 32.

   Pseudonymization does not establish anonymization; legal classifications depend on the actual processing.

8. T01 · rrweb · Accessed: 2026-11-08What is rrweb?
   Passages and scope

   Current documentation; publication date not displayed.

   rrweb architecture: structured events, recording and reconstruction of a web session.

   This web library’s architecture; mobile SDKs may use different methods.

9. T02 · rrweb-io · Accessed: 2026-11-08rrweb ; Incremental snapshots
   Passages and scope

   Project documentation, accessed 8 November 2026.

   Incremental snapshots: mutations, interactions, input changes and scrolling.

   Technical capabilities depend on options, version and integration.

10. T03 · rrweb · Accessed: 2026-11-08Serialization ; rrweb docs
    Passages and scope

    Current documentation; publication date not displayed.

    DOM serialization and reconstruction from retained states and changes.

    Describes rrweb’s choices; reconstructed sessions may be incomplete.

11. T04 · rrweb · Accessed: 2026-11-08Guide ; rrweb: record options, privacy, replay options
    Passages and scope

    Current documentation; publication date not displayed.

    Capture settings, masking and exclusions; playback options.

    An available option does not establish that a site enables it.

12. T06 · Microsoft · Accessed: 2026-11-08Masking content ; Microsoft Clarity
    Passages and scope

    Displayed update: 15 November 2024.

    Clarity: documented pre-transmission masking and protection levels.

    Supplier documentation; setting changes affect future data.

13. T07 · Microsoft · Accessed: 2026-11-08Frequently asked questions ; Microsoft Clarity
    Passages and scope

    Current FAQ, accessed 8 November 2026.

    Clarity is advertised as free; Microsoft states that it does not sell data and acts as a data controller.

    Supplier statements; these do not certify a particular deployment’s compliance.

14. T08 · Fullstory · Accessed: 2026-11-08Fullstory Private by Default
    Passages and scope

    Current help page; publication date not displayed.

    Fullstory: browser-side masking when Private by Default is enabled; distinction from exclusion.

    Settings and content types matter; this documentary reading is not a product test.

15. T09 · Contentsquare · Accessed: 2026-11-08Personal data handling ; Web
    Passages and scope

    Displayed update: 26 October 2026.

    Contentsquare: channel-specific protections and optional encrypted collection, with the customer holding the private key.

    This key arrangement differs from the CNIL’s description of M2.

16. T10 · Princeton CITP; Steven Englehardt, Gunes Acar, Arvind Narayanan · Accessed: 2026-11-08No boundaries: Exfiltration of personal data by session-replay scripts
    Passages and scope

    Published 15 November 2017; historical observations from that study.

    Recording detected on 482 of the 50,000 sites in the selected ranking, reported as a lower bound.

    No measurement of current prevalence; 482 is not a count of password leaks.

17. T11 · USENIX Security 2022; Asuman Senol, Gunes Acar, Mathias Humbert, Frederik Zuiderveen Borgesius · Accessed: 2026-11-08Leaky Forms: A Study of Email and Password Exfiltration Before Form Submission
    Passages and scope

    USENIX Security publication, August 2022; principal data collection May–June 2021.

    The authors identify 52 sites with inadvertent password collection by replay scripts and report remedial changes.

    The email-address findings concern a broader set of tracking technologies.

18. T12 · Asuman Senol et coauteurs; USENIX · Accessed: 2026-11-08Leaky Forms ; presentation slides
    Passages and scope

    Presentation dated 11 August 2022; data from May–June 2021.

    Authors’ slides clarifying the timing and methodology of Leaky Forms.

    Companion to the original paper; historical findings.

19. T13 · Jiawei Guo, Zongze Li, Haipeng Cai; University at Buffalo · Accessed: 2026-11-08TapSpy: Semantic Usage Analysis and Triangulated Risk Assessment of Session Replay Services in Mobile Apps
    Passages and scope

    Public 2026 version; conference scheduled for 15–19 November 2026.

    TapSpy: analysis of Android replay integrations, model-assisted reasoning and partial dynamic validation.

    Classified risks and directly observed leaks are different; no extrapolation to the French web.

20. T14 · ACM SIGSAC · Accessed: 2026-11-08Main Conference Schedule ; ACM CCS 2026
    Passages and scope

    CCS 2026 schedule accessed before the conference.

    Confirms TapSpy’s inclusion in the conference program.

    Inclusion in the program does not mean the work had already been presented on 8 November.

21. T15 · Fullstory · Accessed: 2026-11-08Rage Clicks, Error Clicks, Dead Clicks, and Thrashed Cursor ; Frustration Signals
    Passages and scope

    Current documentation; publication date not displayed.

    Definitions of interaction signals, including repeated clicks within an area.

    A technical signal establishes neither an emotion nor the cause of an exit.

22. T16 · Fullstory · Accessed: 2026-11-08Prevent elements from being classified as Rage or Dead Clicks
    Passages and scope

    Current documentation; publication date not displayed.

    False-positive rage clicks and ways to exclude them; calendar controls provide an example.

    Supplier documentation used to qualify interpretation of the signal.

23. E01 · Contentsquare · Accessed: 2026-11-08Diptyque is transforming its website and business culture with Contentsquare
    Passages and scope

    Publication date and testing calendar not provided.

    Diptyque: reported 9% increase in mobile conversion to address entry, following diagnosis and variant testing.

    Supplier case; starting rates, sample sizes, uncertainty and completed-order outcomes are unpublished.

24. E04 · Fullstory · Accessed: 2026-11-08MOO accelerates issue resolution, increases conversions, and reduces drop offs with Fullstory
    Passages and scope

    Undated commercial PDF; measurement period not published.

    MOO: drop-off levels of 41% and 29%. Our calculation gives −12 percentage points, or −29.3% relative.

    Supplier-reported result; no public protocol or sample sizes allowing reproduction.

25. E05 · MOO · Accessed: 2026-11-08Q&A: How to customize branded merch, with our Head of UX
    Passages and scope

    Published 11 October 2024.

    MOO confirms using Fullstory alongside interviews and customer feedback.

    The customer’s publication does not confirm the supplier case’s numerical gains.

26. E09 · Microsoft Learn · Accessed: 2026-11-08Consent Management ; Microsoft Clarity
    Passages and scope

    Displayed update: 21 May 2026.

    Clarity: Microsoft Ads sharing depends on ad_storage; denied analytics_storage entails limited cookieless collection.

    Documented behavior; no particular site’s activation or revenue per session is established.

27. E10 · Microsoft Learn · Accessed: 2026-11-08Changes to Clarity Reporting Without Cookie Consent
    Passages and scope

    Displayed update: 27 August 2025.

    In the documented mode without cookie consent, each pageview becomes a session and replays cover a single page.

    Read alongside the newer E09. Technical mode and legal consent requirements are distinct.

28. E11 · Microsoft Learn · Accessed: 2026-11-08Data Retention ; Microsoft Clarity
    Passages and scope

    Displayed update: 30 June 2026.

    Clarity’s stated retention: playback 30 days; labeled or favorited sessions 9 months.

    A product retention period does not establish its necessity for a particular purpose.

29. E14 · ACM KDD / Microsoft Research · Accessed: 2026-11-08A Dirty Dozen: Twelve Common Metric Interpretation Pitfalls in Online Controlled Experiments
    Passages and scope

    KDD publication, August 2017.

    Experiment interpretation pitfalls: denominators, population composition and the scope of metrics.

    General methodological research; no measurement of session replay returns.

30. R01 · Data Mining and Knowledge Discovery / Springer · Accessed: 2026-11-08Controlled experiments on the web: survey and practical guide
    Passages and scope

    Published online 30 July 2008; February 2009 journal issue.

    Guide to controlled web experiments: random assignment, statistical power and practical safeguards.

    Methodological support; does not measure the causal return on session replay.