What 883 Real Image Audits Found

Original research · Updated September 15, 2026 · Data collected 2025-02-27 to 2026-09-15

Between February 2025 and August 2026, 883 websites were run through our scanner, which loads a page in a real headless browser and measures every image twice: the pixel size of the file, and the size the browser actually drew it at. This is what those 25,546 images look like in aggregate.

Key findings

The tail is heavier than the median suggests: 37% of measurable images exceed 10x in area, and 63.3% exceed 2x.

35.4% of images have no alt text

This was not what we set out to measure, and it is the finding that surprised us most. Of 21,196 HTML images, 35.4% had a missing or empty alt attribute. An empty alt="" is valid and correct markup for a decorative image, so this figure is an upper bound on the problem rather than a count of defects. But decorative images are a minority of what appears on a typical page, and a rate this high across 883 sites suggests most of it is simply missing.

Format adoption is slower than you would think

By file extension, across all 25,546 images:

Only 12.8% of images use a modern format (WebP plus AVIF), while PNG and JPG together still account for 56.6%. AVIF has barely registered at 0.2%. One important caveat: this counts file extensions, so it undercounts WebP and AVIF delivered by CDNs through content negotiation, where the URL keeps a .jpg extension. The 5% extensionless bucket hides more of the same.

One in five images is a CSS background

4,350 of 25,546 images (17%) were CSS background images rather than <img> elements. This matters methodologically: an audit that only parses HTML for <img> tags misses roughly a fifth of the images on a page, and CSS backgrounds skew toward large decorative hero art — exactly the heavy files an audit most needs to catch.

Median page: 5 images. Heaviest: 1,400.

The median page in the sample carried 6 images, but the mean was 28.9 and the largest single page held 1,400. That gap between median and mean is the whole story of image performance: most pages are modest, and a small number of extreme outliers drag every average upward. Quote the median.

Methodology

• Collection: every URL submitted to the public scanner between 2025-02-27 and 2026-09-15.
• Deduplication: one row per site, using only the most recent scan, so repeat audits of the same URL do not double-count.
• Measurement: pages are loaded in a real headless browser, not parsed as static HTML, so JavaScript-inserted and lazy-loaded images are included. Each image is measured for its intrinsic pixel size and its rendered size.
• "Oversized" definition: intrinsic pixel area divided by rendered pixel area, counted strictly above 4x. Because 4x area is exactly 2x per side, a correctly-served retina image sits precisely on the line and is excluded.
• Measured at 1x: the scanner renders headless at a device pixel ratio of 1, so it compares against CSS pixels. An image served for a 2x retina display therefore appears at about 4x here, which is why the threshold sits exactly there and counts only what exceeds it.
• Measurable subset: 15,506 of 25,546 images had both dimensions available. The rest were hidden, failed to load, or were rendered at zero size.
• Aggregate only: no individual site is named, and no per-site data is published.

Citation

Journalists and researchers are welcome to quote this freely. Suggested form:

ImageDimensions, "What 883 Real Image Audits Found: A Field Study"
(September 2026). Based on 25,546 images across 883 websites
submitted to the ImageDimensions scanner, Feb 2025 - Sep 2026.
https://imagedimensions.com/research/image-audit-field-study-2026

The underlying measurement is reproducible: the scanner that produced this data is free and public, and the CLI and MCP server run the same measurement from a terminal or an AI agent, so any figure here can be checked against your own sites.