16.2% of Images Have No Alt Text. The Fix Costs 5¢ an Image and Can Make It Worse.

Bulk alt text generation runs 2¢ to 5¢ an image. What the tools charge, what they send off your server, and the images they should have been told to skip.

Two rounded square tiles side by side, the left one teal with three short amber bars across it and the right one plain grey and empty
14 min read
2,789 words

WebAIM scanned the top million home pages in February 2026 and found 16.2% of their images carrying no alt text at all. The number underneath it is the one worth your attention: of the images that did have alt text, 10.8% carried values like alt="image", a file name, or a description identical to the sentence sitting beside them. Those are two separate failures, and bulk alt text generation — point a tool at the media library, press the button, pay two to five cents an image — clears the first by quietly feeding the second. The invoice is easy to forecast. What you get for it depends almost entirely on a sorting decision you make before the job runs, and none of the tools will make it for you.

What the backlog actually looks like

WebAIM’s annual scan is the only independent measurement of this at scale, and its 2026 edition covers a million home pages with data collected in February. The average home page now carries 66.6 images — a 13.6% rise in a single year — and 10.8 of them per page have no alternative text. The proportion is improving: 16.2% this year against 18.5% in 2025. It is improving slowly, against a denominator that is growing faster.

The part that turns this from a compliance checkbox into something operational is where the missing alt text sits. WebAIM reports two separate figures here, and both matter. 45% of the images with no alt text were linked images. And, counted the other way round, one out of every four linked images was missing alternative text. A user tabbing through those gets a link with nothing to announce — not a subtle degradation, and concentrated exactly where people are trying to navigate.

WebAIM Million report section stating that 16.2% of home page images had missing alternative text, that 10.8% of images with alternative text had questionable or repetitive values, and that more than one in four images on popular home pages have missing, questionable or repetitive alternative text
The WebAIM Million, 2026 edition, captured 18 Nov 2026: 16.2% of home-page images had no alt text, and a further 10.8% of those that did carried questionable or repetitive values.

WebAIM’s own summary of the two figures together is the sentence to keep in mind for the rest of this: more than one in four images on popular home pages have missing, questionable, or repetitive alternative text.

What bulk alt text generation costs per image

AltText.ai publishes its rate per image rather than making you derive it, which makes it the cleanest benchmark here. On monthly billing, the entry plan is $5 a month for 100 credits — 5¢ an image. It falls to 4¢ at $19 for 500, 3¢ at $59 for 2,000, and 2¢ at both $119 for 5,000 and $229 for 10,000. Annual billing shaves a tier off: $49 a year for 1,200 images at 4¢, or $489 for 24,000 at 2¢. One credit generates one description; 50-credit top-up packs cost $3 with no subscription, and new accounts get 25 credits free.

AltText.ai pricing cards for monthly billing showing Bronze at five dollars a month for 100 credits at five cents per image, Silver nineteen dollars for 500 at four cents, Gold fifty-nine dollars for 2,000 at three cents, Titanium one hundred and nineteen dollars for 5,000 at two cents, and Platinum two hundred and twenty-nine dollars for 10,000 at two cents
AltText.ai's pricing page on monthly billing, captured 18 Nov 2026: $5 a month for 100 images at 5¢ each, falling to 2¢ at $229 a month for 10,000.

Note the shape of that ladder. The rate drops from 5¢ to 2¢ across a hundredfold increase in volume, so there is no cliff where a large library suddenly becomes cheap: a 4,000-image backlog costs roughly $120 at the 3¢ tier, and reaching 2¢ means committing to a volume most sites will never use.

Two surcharges sit outside the headline rate and catch people who optimised their images for speed. AltText.ai charges one credit for standard formats — JPG, PNG, GIF, WebP, BMP — and two for AVIF, HEIC, HEIF, JP2, SVG and TIFF, which also require a paid plan. A site that converted its library to AVIF for Core Web Vitals has doubled its alt text bill without touching a single image. Each additional language costs another credit per image on top, so a bilingual store pays three credits for an AVIF asset it describes twice.

If you already pay for an SEO suite, the arithmetic changes shape rather than disappearing. All in One SEO bills alt text out of a generic AI credit pool at 10 credits per description, against a 10,000-credit pay-as-you-go block listed at $9.99 — so roughly a cent an image, if alt text is all you spend credits on, which it rarely is, since the same pool funds every other AI feature in the plugin. AIOSEO’s credits also expire after a year and do not roll over. That is a different failure mode from a subscription: you are not paying per image so much as racing a clock.

What leaves your server when you press the button

Every hosted option in this category works the same way, and the plugins are commendably direct about it on their own listings. The Alt Text AI plugin — 20,000+ active installations, last updated 22 October 2026 — states that when alt text is generated, your image and your site’s URL are sent to the AltText.ai API. AltTextLab, a much smaller plugin at 100+ installations but updated as recently as 2 November 2026, says it sends the image URL or a base64-encoded copy of the image, along with keywords, the post or product title, and brand attributes depending on how it is configured.

For most sites this is a non-event: a public blog’s photographs are already served to anyone who asks, and a vendor fetching a URL you publish is not an exposure. The cases where it matters are specific, and the bulk button does not pause to ask — a membership site whose media library sits behind a login, a studio holding a client’s unreleased product shots, a clinic or a law firm whose uploads are attached to real people. In those cases the base64 path is the one to look at — it reaches the vendor even when the file itself was never publicly addressable, which is precisely the protection you thought you had.

It is the same calculation as any other outsourced inference, and it resolves the way it does when you ask whether your business actually needs a local LLM: for public content, send it; for anything else, spend ten minutes on it before you queue four thousand uploads.

The local route, and the eighteen-month hole in the evidence

The obvious answer to both the cost and the privacy question is to run the model yourself, and the tooling for it is genuinely cheap now. Ollama’s current library lists lfm3-vl:2b, a vision-language model built for on-device use: 1.92 billion parameters in total, split across a 1.73B language model, a 164M vision encoder and a 25M projector, for a download of 1.5 to 1.7GB. It processes images at native resolution. At the other end of the same library sits qwen3.8:27b at 18GB with a 256K context window and 6.8 million downloads, the most-pulled vision-capable model Ollama lists. Neither charges per image. The first cost is the download, which — as with every local model whose download size has nothing to do with its parameter count — is the number to check before the parameter count.

Then the honest part. The only published, scored comparison of local vision models against cloud ones for this specific job is Dries Buytaert’s, and his most recent round is dated 27 May 2025. It is worth knowing the shape of it before leaning on the numbers: five images, scored by hand, on a scale he devised himself, and he says plainly that the set is small and the scoring subjective. He marked each description from 0 to 5 on whether it caught the important objects, whether it conveyed mood, and whether it repeated itself or hallucinated. Qwen 2.5 VL (32B) came out best of the local models at 4.4, Gemma 3 (27B) at 4.2, Mistral 3.1 (24B) at 3.4 — against the cloud scores from his earlier round, GPT-4o at 4.8 and Claude 3.5 Sonnet at a clean 5. None of the three local models hallucinated or repeated itself, which is the failure that makes older captioning models unusable for screen readers. What they did instead was miss things: no local model identified Shibuya Crossing, all three called wakeboarding “water skiing”, and only one of them recognised a backgammon board as a backgammon board.

Buytaert’s own conclusion was that the quality was “not yet at the level where I would trust these models” to generate thousands of descriptions without human review. He is no sceptic: he says plainly that he prefers local models for privacy, cost and offline use, and he had about 9,000 images on his own site waiting. He used a cloud model for them anyway.

Here is what matters for your decision: that test is eighteen months old, and it tested models that are no longer the ones you would run. Gemma 3, Qwen 2.5 VL and Mistral 3.1 were all released in March 2025. None of them appears on Ollama’s vision listing today. Nobody has rerun the comparison on the models that replaced them. So the correct statement about local alt text generation in November 2026 is not that it has caught up, and not that it hasn’t — it is that the gap is unmeasured. If you go local, you are running a sample and judging it yourself, which is work, but it is the same work the cloud route needs and nobody does. If you want the ground floor of that setup, the complete guide to running local LLMs with Ollama covers the install, and what LFM3’s tiny models actually deliver covers what to expect from the 2B end of the range.

The decision no generator makes for you

All of the above is a question about model quality, which is the wrong thing to spend your attention on. The decision that determines whether a bulk run helps or hurts does not depend on the model at all.

W3C publishes a decision tree for choosing alt text, and several of its branches end in an empty alt attribute rather than a description. Two of them are the ones a bulk run walks straight into: an image that is “purely decorative or not intended for users”, and an image whose content is redundant to real text already sitting next to it.

W3C alt decision tree showing the branch for images that contribute meaning, where content redundant to real text nearby takes an empty alt attribute, and the branch asking whether an image is purely decorative, where the answer yes takes an empty alt attribute
W3C's alt Decision Tree, captured 18 Nov 2026. Two of its branches end in an empty alt attribute, and both of them depend on the page around the image rather than the image itself.

Both of those are judgements about the page, not about the picture. A generator is handed an image URL or a base64 blob. It cannot see the heading above the image, the caption below it, or the author’s name printed next to the author’s photograph. So it describes the section divider. It describes the background texture. It writes “a smiling woman with shoulder-length brown hair” under the byline that already says who she is, and it writes a careful sentence about the product on the product page where the caption underneath says the same thing in better words.

Go back to WebAIM’s second statistic and read what it actually counts. Some of that 10.8% is ordinary human neglect — alt="image", alt="graphic", a file name left as-is. But one of the things it counts is alt text “identical to adjacent text or the alternative text of an adjacent image”, and that is not careless human writing. It is a precise description of what a context-blind generator produces at scale, already measurable across the top million home pages before most of them have run one.

There is a default in this category that gets the rule exactly backwards. AltTextLab, to its credit, protects existing alt text: it only fills in images that currently have an empty alt attribute. Read that against the decision tree and the problem is immediate — an empty alt is the correct, deliberate, standards-compliant value for a decorative image. A site that has done its accessibility work properly has marked its dividers and spacers with alt="", and those are the first images the plugin overwrites.

A bulk run that is worth the money

The sequence matters more than the tool, and most of it happens before anything is generated.

  1. Count and cost it first. Multiply your library against the published per-image rate at the tier you would actually buy, and check your formats — if you converted to AVIF or WebP for performance, confirm which your vendor bills at two credits rather than one.
  2. Classify before you generate. Split the library into images that carry meaning and images that do not. Decorative graphics, dividers, background textures, spacer images, and any image whose caption or adjacent heading already says the same thing get alt="" and are excluded from the job entirely. This is the step that decides the outcome, and it is the only one no tool performs.
  3. Settle the privacy question once, on what the plugin’s own listing says it transmits rather than what the marketing page implies. If any of the library sits behind a login, that is a decision, not a default.
  4. Run a sample, not the library. AltText.ai’s free tier is 25 credits and AltTextLab’s is 25 a month — enough to see the register a tool writes in, which the sales page cannot tell you.
  5. Read the sample against the pages. The failures Buytaert catalogued are specific and recognisable once you know to look: a landmark left unnamed, an activity misidentified, a description that is accurate about the picture and useless on the page it sits on.
  6. Script the remainder where the tool supports it. The Alt Text AI plugin exposes wp alttext generate and wp alttext status, which turns a four-thousand-image job into batches you can check between runs rather than a browser tab you must not close.
  7. Spot-check the linked images first. WebAIM found 45% of missing-alt images were links. On a link, a wrong description is not a cosmetic flaw — it is a wrong label on a control, and the user acts on it.

Where this still goes wrong

Charts, graphs and screenshots are a category error rather than a quality problem. W3C routes complex images to an explanation elsewhere on the page, not to an alt attribute, so a generator producing a beautiful sentence about your revenue chart has failed the rule however good the sentence is.

Context is the other permanent limit. The same photograph of a warehouse needs different alt text on a careers page, in a news story about a fire, and on a product listing. The model sees the warehouse; it cannot see which of those pages it is on, and no increase in model quality changes that.

The cost nobody budgets is the review. Two cents an image is not the expense at four thousand images; the hours someone spends reading four thousand generated sentences is. It is also the cost most people quietly skip, which is how the output lands in WebAIM’s second bucket instead of fixing the first. Credit models add their own pressure here — AIOSEO’s expire after a year with no roll-over, which rewards generating rather than checking.

And the honest ceiling: automated alt text raises a site’s floor. It does not reach what an editor writing with the page in front of them produces, and the WebAIM trend line suggests the problem is receding year on year without anyone’s bulk button. What the button buys you is speed on the images that genuinely need describing — which is a smaller set than your media library, and knowing which ones they are is the entire job.

Sources

Every figure in this article traces to one of these. Dates are when the page was read.

  1. The WebAIM Million — The 2026 report on the accessibility of the top 1,000,000 home pages
    WebAIM, Utah State UniversityAccessed Nov 18, 2026Independent
  2. An alt Decision Tree
    W3C Web Accessibility Initiative (WAI)Accessed Nov 18, 2026Official docs
  3. Comparing local LLMs for alt-text generation, round 2
    Dries BuytaertAccessed Nov 18, 2026Independent
  4. Comparing local large language models for alt-text generation
    Dries BuytaertAccessed Nov 18, 2026Independent
  5. AltText.ai Pricing
    AltText.aiAccessed Nov 18, 2026Vendor pricing
  6. Alt Text AI — Automatically generate image alt text for SEO and accessibility
    WordPress.org Plugin DirectoryAccessed Nov 18, 2026Official docs
  7. AI Alt Text Generator for Images — AltTextLab
    WordPress.org Plugin DirectoryAccessed Nov 18, 2026Official docs
  8. Generating Image Alt Text Using AIOSEO AI
    All in One SEOAccessed Nov 18, 2026Vendor pricing
  9. lfm3-vl
    Ollama model libraryAccessed Nov 18, 2026Official docs
  10. qwen3.8
    Ollama model libraryAccessed Nov 18, 2026Official docs

Written by

Kai Devlin AI Automation Engineer

Full-stack developer who builds and runs AI automation systems in production. Runs local LLMs on personal hardware, builds N8N pipelines that actually ship, and deploys on Cloudflare Pages. Every guide on Pipeline Monk is tested on real consumer hardware — a Ryzen 7 5800HS with 16GB RAM and a GTX 1650. If it works on that, it works.