Updated October 2026: corrected Netflix Dynamic Optimizer and Bitmovin savings figures, added a technique-by-technique comparison table, 2025–2026 changes (AV1 at 30% of Netflix streaming, AV2, NETINT 2026 survey, Brightcove and Azure vendor changes) and a note on the unrelated security term "context-aware output encoding".

Why this matters

If your team pays a content delivery network bill, content-aware encoding is the cheapest large dollar-saving you can make in your video pipeline this year. The savings are concentrated in your most-watched titles, the engineering work is incremental rather than a full re-platform, and the technology is mature enough that every major cloud encoder offers a turnkey version. Context-aware encoding is the second lever — it stops you from sending a 4K rung to a viewer who is decoding on a 720p phone screen on a 4G connection, which is the silent waste in most "we already encode in HD/SD/4K" pipelines.

This article is for product managers, founders, video operations leads, and engineering managers who need to read a vendor pitch — "AI content-aware encoding, 40% bitrate savings" — and know which lever the vendor is pulling, what is real, and what to ask next. We start from the simplest version of "encode each video to its own complexity", walk through how that idea expanded into per-scene, then into region-of-interest, then into delivery-context optimisation, and finish with a decision tree for which approach to deploy first.

Two words, two meanings — and the industry has been sloppy

The phrase "content-aware encoding" is older than "context-aware encoding" by about two years (Netflix per-title in 2015, Brightcove context-aware in 2017), and the industry has used both terms loosely ever since. Before going further, lock down the definitions, because every "AI encoder" pitch you will read this year mixes them.

Content-aware encoding looks at properties of the video itself. Cartoon or sports? Talking head or aerial drone shot? Dark interior or bright exterior? Flat sky or grass texture? The encoder uses those properties to decide how many bits to spend on the whole title, on each scene, or even on each region inside each frame. The technology is also called content-adaptive encoding (CAE), per-title encoding, per-scene encoding, shot-based encoding, region-of-interest encoding, and content-aware bitrate ladder generation. Those are not synonyms — they sit at different levels of granularity — but they all share the same definition: the input is the source video, the output is a bitrate-allocation decision driven by what the encoder sees.

Context-aware encoding looks at properties of delivery — who will receive the video, on what device, over what network. A phone on 4G with a 6-inch 720p screen does not need a 4K rung; a 65-inch TV on fibre does not need a 360p rung. The technology was named by Yuriy Reznik at Brightcove in 2017, and the cleanest published definition is in his DASH-IF 2019 workshop slides: context-aware encoding "exploits diversity of properties of delivery devices… and access networks", with the objective of "generating a multiset of encoded sequences, which can be optimally used for ABR streaming to devices and access networks of each kind". 6 The input is the device-and-network mix of your audience; the output is the adaptive bitrate ladder — the set of resolution and bitrate rungs your player can switch between during playback.

The two are complementary. Content-aware encoding tells you, for this video, what the optimal rate-quality curve looks like. Context-aware encoding tells you, for this audience, which rungs along that curve you should actually ship. In production, leading streaming platforms apply both, often inside a single pipeline. The 2018 Streaming Media "Buyers' Guide" was the first piece to name this distinction in writing; eight years later most marketing pages still smudge it together as "CAE", which is why product teams routinely buy one when they wanted the other. 8

In plain terms, "content-aware" means the software looks at the content itself before deciding what to do with it. One search-term trap: context-aware output encoding is a web-security practice, not a video technique. It means escaping untrusted data for the HTML, JavaScript, URL or CSS context it is written into, and OWASP's Application Security Verification Standard lists it as a defence against reflected, stored and DOM-based cross-site scripting. 20 If that is what you are looking for, nothing in this article applies.

Two panels side by side. Left, in blue, Content-aware encoding, which looks at the video. Its input is source video frames, per-shot complexity, and saliency and face regions; an arrow leads down to its output: a per-title ladder for 20 to 50% less bandwidth, a per-shot bitrate for a further 27 to 37% as measured by Netflix, and a per-region QP for region-of-interest coding, worth 3 to 5 dB PSNR on faces. Right, in navy, Context-aware encoding, which looks at the audience. Its input is the device class mix, the bandwidth distribution as a CDF, and codec support per device; its output is a rung count of 4 to 6 instead of 10 to 12 for 30 to 50% less storage, rungs placed near the audience modes for a higher average viewer VMAF, and a codec choice per cohort for a better fit to each device. A closing line says content-aware shapes the rate-quality curve of each video, and context-aware picks which rungs of that curve you actually ship. Figure 1. The two ideas share an acronym but solve different problems. Content-aware encoding decides how many bits to spend on each part of the source video. Context-aware encoding decides which rungs of the resulting bitrate ladder to actually ship to your audience, given the devices and networks they use.

Content-aware encoding — three levels of granularity

Content-aware encoding has three working depths, and you should know all three because the savings, the engineering cost, and the operational risk scale with the depth.

The shallowest is per-title encoding, where the pipeline picks a single bitrate ladder for the whole video based on its global complexity. Netflix published the first widely-cited per-title paper in December 2015 and triggered the industry-wide adoption of content-aware techniques. 1 The recipe: encode the title at several resolutions and several quality settings to produce a cloud of (resolution, bitrate, quality) points; draw the convex hull (the outer envelope of the best-of-best points) on a rate-quality plot; place ladder rungs along that envelope at the bitrates your audience actually uses. The quality score is VMAF (Video Multi-method Assessment Fusion), Netflix's open-sourced perceptual metric that scores 0–100, where higher means closer to the source. 4 Per-title typically saves 20–50% bandwidth at the same VMAF.

The middle is per-scene or shot-based encoding, where the pipeline picks a fresh bitrate for every shot. A shot is one continuous camera capture, and a typical 90-minute film contains 700–1,200 shots at an average of about four seconds each. Netflix's Dynamic Optimizer, deployed in production in 2018, encodes each shot at several quality points, runs a Lagrangian optimisation across all shots in the title — picking the per-shot quality that minimises total bits for a target average VMAF — and stitches the result back together. 2 3 Netflix reported that the framework improves H.264, HEVC and VP9 encodes by roughly 28–38% at the same quality (Streaming Media later summarised it as 28% for x264, 32% for HEVC and 37% for VP9 19), and, in its most conservative test, 17.1% average savings for VP9 over the best possible fixed-QP encode of the whole title. 2 Bitmovin's commercial Per-Scene Adaptation claims up to 30% bandwidth savings; in its own 2018 test clips the gain ranged from about 3% on sports to 14% on animation and 28% on nature footage, which is a useful reminder that the saving depends on the content. Mux, AWS Elemental, Brightcove, Harmonic, and Ateme all ship their own versions. 5

The deepest is region-of-interest (ROI) encoding, where the encoder varies the quantization parameter (the knob that controls how aggressively each block is compressed) inside a single frame. The classical use case is video conferencing: spend more bits on the face, fewer on the background. A face-detector — modern pipelines use a small CNN like RetinaFace or MediaPipe Face Detector running once every few frames — labels the face region, and the encoder lowers the QP inside that region and raises it outside. Published academic ROI implementations report 3–5 dB PSNR improvement and 15–20% SSIM improvement in the face region at the same total bitrate. 10 11 ROI encoding ships today in surveillance cameras such as Hikvision's, where it is a standard encoder setting. It is the deepest, most expensive form of content-aware encoding because the encoder must run the detector and feed its output into the rate-distortion loop, but for verticals where the viewer's eye is reliably on one region — faces in a call, license plates in surveillance, the puck in hockey — the perceptual gain is large.

These three depths nest. Per-scene encoding builds on top of per-title (you still need a sensible per-title ladder to know what bitrates to target); ROI encoding builds on top of per-scene (you apply ROI passes inside the per-shot encodes). A mature 2026 pipeline runs all three: it picks a per-title ladder, splits the title into shots, runs a Lagrangian optimisation over the shots, and then applies ROI inside each shot for verticals where the face/object-of-interest gain pays for the detector compute.

Three nested boxes. The outer box, level 1, per-title encoding: one ladder for the whole title, 20 to 50% less bandwidth than a fixed ladder. Inside it, level 2, per-scene or shot-based encoding: about 900 shots in a film, each at its own bitrate, for a further 27 to 37%. A row of sixteen bars from shot 1 to shot 900 shows the bitrate rising and falling from shot to shot; one bar is green, and a dashed line runs from it down into the innermost box. Level 3, region of interest: inside one shot, lower QP on the face and higher QP on the background, for 3 to 5 dB more PSNR on the face at the same bitrate. Beside it a grey frame labelled background, high QP, holds a green oval labelled face, low QP. A closing line says per-title sets the budget, per-scene moves bits between shots, ROI moves bits inside the frame, and mature pipelines run all three. Figure 2. The three depths of content-aware encoding nest inside each other. Per-title sets the overall budget. Per-scene redistributes the budget across shots. ROI redistributes bits inside each frame toward the regions the viewer looks at.

Show me the math — per-scene savings in one paragraph

Take a 90-minute drama with a per-title 1080p bitrate of 4 Mbps. The total payload for one viewer is:

size = bitrate × time
     = 4 Mbps × 5,400 seconds
     = 21,600 megabits
     = 2,700 megabytes
     ≈ 2.70 GB

Now split the title into 900 shots and assume the per-scene optimiser saves 17% on average, close to the 17.1% Netflix reported in the most conservative Dynamic Optimizer test. 2 The new payload is:

size_new = 4 Mbps × (1 − 0.17) × 5,400 s
         = 3.32 Mbps × 5,400 s
         = 17,928 Mb
         = 2,241 MB
         ≈ 2.24 GB

That is a 459 MB per-view saving. At one million views per year and a CDN price of one cent per gigabyte, the saved bandwidth is worth:

savings = 0.459 GB × 1,000,000 × $0.01
        = $4,590 per year per title

The arithmetic — per-view delta × view count × unit cost — is the entire business case in one line. The percentage gain looks small on a single title; the dollar value scales with viewership and concentrates in the catalogue's most-watched titles. For a streaming service with 1,000 active titles and a long-tail viewership distribution, the savings stack to mid-six-figures annually for an engineering investment that is mostly one quarter of pipeline work.

Context-aware encoding — encoding for the audience you actually have

Context-aware encoding starts from a question that fixed bitrate ladders never ask: who is going to watch this rung, and on what? A 4K rung is wasted on a viewer whose phone has a 720p screen and a 4G connection. A 360p rung is wasted on a viewer with a 65-inch TV and gigabit fibre. The Apple HLS Authoring Specification's default ladder ships rungs from 145 kbps at 416×234 up to 7.8 Mbps at 1920×1080, and most pipelines copy them verbatim regardless of who their audience is. 9 Context-aware encoding throws out the fixed list and computes a ladder tailored to your viewers' actual devices and networks.

Reznik's 2017 Brightcove formulation, which is still the cleanest in the literature, models the audience as a joint distribution over device class (mobile, tablet, PC, connected TV), screen resolution, and access-network bandwidth (the empirical CDF of bandwidth across your viewers). The ladder is then designed as an optimisation problem: minimise the expected distortion over that distribution, subject to a constraint on the number of rungs and the maximum bitrate. The solution typically has 4–7 rungs (not the 9–12 of the Apple default), with rung spacing that hugs the modes of the viewer-bandwidth distribution and resolutions that match the modal screen sizes. 6 7

The user-visible effect is twofold. First, for the same total storage budget, viewers see better quality on average because the rungs are placed where the audience actually sits on the bandwidth distribution. Second, for the same average quality, total storage drops because redundant rungs are removed — a typical context-aware pipeline ships 4–6 rungs instead of 10–12, cutting storage by 30–50% before any per-title savings are applied. The two saves compose: if per-title saves 30% per rung and context-aware halves the number of rungs, total storage falls by 60–70% versus a fixed-ladder baseline. 7

The pitfall here is the easiest to hit. Context-aware encoding is sensitive to the audience distribution you measure, and the distribution drifts. If you build a ladder against a 2022 audience that was 60% mobile, then your audience shifts to 40% mobile and 30% TV by 2026, your ladder is now wrong for the people you actually have. The fix is operational: re-measure the device-and-network mix at least quarterly, re-run the optimisation, and stage the new ladder behind a feature flag so you can roll back if the player picks up the wrong rung for a measurable cohort. 7

Two stacked plots of quality against bandwidth with the same rising, flattening rate-quality curve. Top: the fixed Apple HLS default ladder, ten orange rungs spaced evenly along the curve. Bottom: a context-aware ladder of five green rungs tuned to the audience, drawn over a blue histogram of the share of viewers at each bandwidth. The histogram has a large low-bandwidth hump, a larger mid-bandwidth hump and a small high-bandwidth one; one rung sits on the first hump, two on the second, and two near the top. A closing line says fewer rungs where viewers sit give 30 to 50% less storage and better average quality. Figure 3. A fixed ladder spaces rungs evenly along the rate-quality curve. A context-aware ladder packs rungs near the modes of the audience-bandwidth distribution and skips bitrates that no viewer can actually consume. The total storage drops; the average viewer's perceived quality rises.

Common mistake: confusing content-aware with context-aware in a vendor pitch

A specific pitfall every product team hits when buying or building an encoding pipeline: a vendor pitch says "CAE — 40% bandwidth savings" and the team does not ask which CAE. Content-aware and context-aware solve different problems and stack on top of each other; if you buy one thinking it is the other, you do not get the savings you signed up for.

The simple test, drawn from the Reznik framework: ask the vendor what their tool's input is and what its output is. A content-aware tool takes the source video and outputs an encoding plan (bitrate per scene, QP per region). A context-aware tool takes audience statistics and outputs a ladder design (rung resolutions and bitrates). If the vendor cannot answer that question directly — or if the answer is "we use AI" — assume the product is one of the two and ask which.

The four techniques side by side:

TechniqueDecidesInputOutputFirst shipped at scale
Per-title (content-aware)Bitrate ladder for the whole titleSource videoOne ladder per titleNetflix, December 2015
Per-scene / shot-based (content-aware)Quality and bitrate per shotSource video, split into shotsPer-shot encode settingsNetflix Dynamic Optimizer, 2018
Region of interest (content-aware)QP inside each frameSource video plus a face or object detectorQP map per frameSurveillance cameras and conferencing encoders
Context-awareWhich rungs to shipAudience device and bandwidth statisticsLadder design (rung count, resolutions, bitrates)Brightcove, 2017

A second tell: published numbers. Content-aware encoders publish results in BD-rate or VMAF-at-bitrate terms ("17% bitrate reduction at the same VMAF"). Context-aware encoders publish results in rungs-removed and expected-distortion-over-audience terms ("ladder cut from 11 rungs to 5, average viewer VMAF up 1.8 points"). A vendor that publishes neither is a vendor whose technology is not yet measured against its own claim.

Where the AI fits — and where it does not

Both content-aware and context-aware encoding existed before the modern deep-learning wave. Netflix's 2015 per-title was a brute-force search over resolution and bitrate; Reznik's context-aware was a mathematical optimisation over a measured audience distribution. Neither required a neural network. In 2026 the picture is more nuanced, and the marketing has run ahead of the engineering.

The places AI genuinely earns its keep in content-aware encoding are three. First, content-complexity prediction — instead of brute-forcing every resolution and bitrate combination, a small CNN predicts the convex hull from a single fast analysis pass, cutting the encode time of per-title from "encode the whole catalogue ten times" to "encode the whole catalogue once plus a few seconds of analysis per title". 12 Second, saliency-driven ROI — a saliency CNN trained on eye-tracking data predicts which regions of a frame viewers will look at, and the encoder bumps bit allocation there. Third, shot-boundary detection — a small classifier outperforms the classical frame-difference heuristic on fades, dissolves, and high-motion shots, which is where per-scene optimisation gains the most. We covered the first and third of these in detail in AI Inside the Encoder.

The place AI is mostly marketing, today, is "AI-driven context-aware encoding". The audience distribution is well-modelled by classical statistics; the ladder optimisation is a convex problem that any solver handles in milliseconds. A vendor pitching a neural network to do this is solving a problem that does not need a neural network. The honest version of "AI context-aware" is "we use ML to measure the audience distribution more accurately" — clustering viewer cohorts, predicting tomorrow's mix from today's logs — and the gains from that are real but small. If a vendor claims a 40% ladder-storage win from AI alone, the headline is almost certainly content-aware, not context-aware, hiding inside a context-aware label.

What changed for content-aware encoding in 2025 and 2026

The method is the one Netflix and Brightcove described years ago. The codec mix underneath it and the vendors selling it have moved.

AV1 is now a mainstream rung. Netflix said in December 2025 that AV1 carries about 30% of its streaming. 15 A convex hull belongs to one codec, so each codec you add needs its own per-title or per-shot analysis: an H.264 ladder does not carry over to AV1. AV2 is the next one to plan for. AOMedia finalised the AV2 specification in late 2025, with measured gains over AV1 in the high-20% to mid-30% range, and as of 2026 there is no decoder silicon timeline. 16

Adoption keeps growing. NETINT's 2026 State of Video Encoding survey of 286 video professionals (July 2026) found AV1 in production at 17% of organisations and reported 77% planned growth for content-aware ladder generation. 14

The vendor map changed too. Bending Spoons closed its $233 million acquisition of Brightcove, the company that named context-aware encoding, in February 2025. 17 Microsoft retired Azure Media Services, along with its content-aware encoding preset, on 30 June 2024. 18 When the ladder logic lives only inside one vendor's preset, a vendor change means rebuilding it, so keep your audience measurements and quality targets in your own pipeline.

How to decide which to add first

The decision tree, drawn from how mature streaming platforms actually deploy these technologies:

If your pipeline still ships a copy of the Apple HLS default ladder and you have not measured content complexity, start with per-title. The engineering cost is one to two engineer-quarters; the dollar savings concentrate in your top 10% of titles by viewership; the technology is offered as a turnkey feature by every major cloud encoder (AWS Elemental's Automated ABR, Bitmovin's per-title, Mux's automatic ladder). 13 Skip to per-scene only after per-title is in production and stable, because per-scene needs reliable shot detection, an optimisation framework, and a quality model — and most of the catalogue gain is already captured by per-title.

If you have per-title in production, add per-scene next if your catalogue contains content with large within-title complexity swings (sports, action, anything with cuts between calm and busy scenes). For talking-head content, the gain is small enough that the engineering cost is not worth it.

If your service is conferencing, surveillance, or any vertical where the eye reliably lands on a known region, add ROI before or instead of per-scene. The face-detection or object-detection model is the long pole, but most pipelines already have one for other reasons (mute-when-not-looking, background blur, motion alerts).

Layer context-aware on top once your audience is large enough that the device-network distribution is statistically meaningful — practically, once you have more than around 100,000 monthly viewers. Below that, the variance in your audience-mix measurement is larger than the gain from rung-tailoring. Above that, a 4–6-rung context-aware ladder versus a 10–12-rung fixed default is a 30–50% storage saving for one engineer-quarter of work.

Finally, measure with the right metric. Content-aware and ROI both deliberately shift bits away from regions the eye does not notice — they will raise PSNR-measured distortion on purpose. Benchmark with VMAF or SSIMULACRA2, not PSNR. See Objective Quality Metrics for why these metrics exist and when to use which.

A decision tree of four questions stacked top to bottom, each with a yes arrow to a green action on its right; after each question you move down to the next one either way. Question 1, still on the Apple HLS default, the same rungs for every title? Yes: add per-title, 20 to 50% less bandwidth at the same VMAF, one to two engineer-quarters of work. Question 2, complexity swings within a title, such as sports, action, calm-to-busy cuts? Yes: add per-scene, a further 27 to 37% as measured by Netflix, which needs reliable shot detection. Question 3, is the eye reliably on a known region, such as faces, plates, doorways or the puck? Yes: add ROI encoding, 3 to 5 dB PSNR on the face, reusing a detector you already run. Question 4, over 100,000 monthly viewers, so the audience mix is measurable? Yes: add a context-aware ladder, 4 to 6 rungs instead of 10 to 12 and 30 to 50% less storage. Closing lines: this is the order mature streaming platforms follow; benchmark each step with VMAF or SSIMULACRA2, never PSNR. Figure 4. The decision tree mature streaming platforms follow when adding content-aware and context-aware techniques to their pipeline. Per-title first; per-scene if complexity swings inside a title are large; ROI for face- or object-centric verticals; context-aware once audience size makes the distribution measurable.

↓ Free download: Content-Aware Encoding PlaybookOne page: the decision tree for which technique to add first, a ladder-design checklist, and the vendor-question script that tells content-aware from context-aware in a pitch.

Where Fora Soft fits in

Content-aware and context-aware encoding cut across most of what we build. Per-title and per-scene pay back fastest in catalogue streaming, so they are part of the job when our OTT and IPTV platform development team builds VOD libraries and multi-device apps. For live and on-demand pipelines outside OTT, our video and audio streaming software development work covers ladder design for HLS and DASH, measured against the real device and bandwidth mix of your audience. ROI encoding belongs where the eye is predictable: faces in video calling and conferencing apps and in telemedicine consults, where the gain matters most at the low bitrates consumer connections run at, and motion-tracked regions, license plates and doorways in video surveillance software. We have shipped pipelines that combine all of these on top of FFmpeg, GStreamer, x265, SVT-AV1, and the major cloud encoders.

Video Encoding: A Complete Guide to Video Compression - book cover

The book

Video Encoding: A Complete Guide to Video Compression — From Pixels and Color to H.264, HEVC, AV1, Quality Metrics, and AI

Volume 1 takes the rate-quality curve apart: how rate control and QP turn a bitrate budget into pictures, why a convex hull of encodes beats any fixed ladder, and how VMAF and BD-rate let you check a “40% savings” claim instead of taking it on trust. It is the ground floor under per-title, per-scene and ROI encoding, so you can size each lever for your own catalogue before you commit a quarter of engineering to it. Written by Nikolay Sapunov, CEO at Fora Soft.

Read it on Amazon →Kindle & paperback · Fora Soft Video Engineering Handbook

Which lever would cut your CDN bill first: per-title, per-scene, ROI or a new ladder?

Send us your current ladder, a rough view count per title and your device mix — we will tell you which technique pays back first, what it should save in bandwidth and storage, and what to ask a vendor who promises 40%. 250+ video products shipped since 2005.

Book a 30-min call →WhatsApp →Email us →

See our case studies


  1. Netflix Technology Blog, "Per-Title Encode Optimization", 2015. https://netflixtechblog.com/per-title-encode-optimization-7e99442b62a2 ↩

  2. Netflix Technology Blog, "Dynamic Optimizer — a perceptual video encoding optimization framework", 2018. https://netflixtechblog.com/dynamic-optimizer-a-perceptual-video-encoding-optimization-framework-e19f1e3a277f ↩↩↩

  3. Netflix Technology Blog, "Optimized shot-based encodes: Now Streaming!", 2018. https://medium.com/netflix-techblog/optimized-shot-based-encodes-now-streaming-4b9464204830 ↩

  4. Netflix, "VMAF: Perceptual video quality assessment based on multi-method fusion." https://github.com/Netflix/vmaf ↩

  5. Bitmovin, "Per-Scene Adaptation: Going Beyond Bitrate". https://bitmovin.com/per-scene-adaptation-going-beyond-bitrate/ ↩

  6. Y. Reznik, "Context-Aware Encoding and 5G", DASH-IF Workshop, 2019. https://dashif.org/docs/workshop-2019/03-Reznik-DASH-IF-workshop-2019-CAE.pdf ↩↩

  7. Y. Reznik et al., "Improving Context-Aware Encoding by Adaptation to True Bandwidth Distribution", SMPTE Media Technology Summit, 2023. https://www.reznik.org/papers/SMPTE_MTS2023_Improving_Context_Aware_Encoding.pdf ↩↩↩

  8. J. Ozer, "Buyers' Guide to Content- and Context-Aware Encoding 2018", Streaming Media Global. https://www.streamingmediaglobal.com/Articles/Editorial/Featured-Articles/Buyers-Guide-to-Content--and-Context-Aware-Encoding-2018-123407.aspx ↩

  9. Apple Inc., "HLS Authoring Specification for Apple Devices". https://developer.apple.com/documentation/http-live-streaming/hls-authoring-specification-for-apple-devices ↩

  10. MDPI Applied Sciences, "Region-of-Interest Based Coding Scheme for Live Videos", 2024. https://www.mdpi.com/2076-3417/14/9/3823 ↩

  11. Hikvision, "ROI (Region of Interest) Encoding Technology V0.5 White Paper". https://www.hikvision.com/content/dam/hikvision/usa/white-papers/hikvision_roi_encoding_technology_v0.5.pdf ↩

  12. H. Amirpour et al., "Towards Sustainable Video Streaming: Evaluation of AI Approaches for Content-Aware Encoding", IEEE 2024. https://ieeexplore.ieee.org/document/10389538/ ↩

  13. AWS Elemental, "Automated ABR Configuration in AWS Elemental MediaConvert". https://aws.amazon.com/blogs/media/introducing-automated-abr-adaptive-bit-rate-configuration-a-better-way-to-encode-vod-content-using-aws-elemental-mediaconvert/ ↩

  14. NETINT, "2026 State of Video Encoding". https://netint.com/2026-state-of-video-encoding/ ↩

  15. Netflix Technology Blog, "AV1 — Now Powering 30% of Netflix Streaming", December 2025; reported by TV Technology. https://www.tvtechnology.com/news/av1-open-video-codec-now-powers-30-percent-of-netflix-streaming ↩

  16. J. Ozer, "The State of Streaming Codecs 2026", Streaming Media. https://www.streamingmedia.com/Articles/Editorial/Featured-Articles/The-State-of-Streaming-Codecs-2026-173838.aspx ↩

  17. Business Wire, "Bending Spoons closes $233 million acquisition of Brightcove", 4 February 2025. https://www.businesswire.com/news/home/20250204502753/fr ↩

  18. Microsoft Learn, Azure Media Services retirement notice ("Azure Media Services will be retired June 30th, 2024"). https://learn.microsoft.com/en-us/azure/media-services/latest/encode-concept-preset-overrides ↩

  19. J. Ozer, "The Case for Shot-Based Encoding", Streaming Media, 2021. https://www.streamingmedia.com/Articles/Editorial/Short-Cuts/The-Case-for-Shot-Based-Encoding-145473.aspx ↩

  20. OWASP, Application Security Verification Standard 4.0, V5.3 "Output Encoding and Injection Prevention". https://owasp-aasvs4.readthedocs.io/en/latest/V5.3.html ↩