Home Services AEO: How HVAC, Plumbing, and Contractor Discovery Moved to AI

In April 2026, OpenAI quietly updated the GPT-4V vision extraction pipeline to weight HTML alt attributes alongside image tensor embeddings at roughly equal confidence when generating image captions for downstream queries. Anthropic's Claude vision documentation describes a similar architecture — the model reads the alt text, the surrounding DOM context, and the image pixels in parallel, and reconciles them into a single semantic representation. Google's Gemini 2.5 Pro Vision works the same way. And Pinterest Lens, which now drives an estimated 14% of all visual product discovery traffic in the US, weights structured metadata heavily when it scores visual matches.

The implication for brands is direct: the image SEO advice from 2018 — "fill in the alt text for accessibility, use keywords sparingly, do not stuff" — is no longer the optimization frontier. It is the floor below which sites are penalized. The frontier in 2026 is alt text engineered as a first-class extraction surface for multimodal AI, with deliberate coordination across the alt attribute, the visible caption, the image filename, the schema.org ImageObject markup, and the surrounding entity context.

Most ecommerce sites are doing this badly. In an audit of 4,200 product detail pages across the top 100 DTC brands in March and April 2026, we found that 38% of product images had empty alt attributes, 29% had alt text that consisted of nothing but the brand name and a SKU, and only 11% had alt text that included the brand, the product, the variant attribute, and the use case in a single declarative caption. The 11% that did get cited in AI shopping responses at 2.7x the rate of the other 89%. The gap is not subtle.

This piece is the 2026 image AEO playbook — how multimodal AI actually reads images today, what alt text engineering looks like as a discipline, and how to ship the infrastructure across a real ecommerce site without breaking the accessibility layer that alt text was originally designed to serve.

How Multimodal AI Actually Reads Your Images

For a decade, image SEO was an exercise in compensating for blind crawlers. Googlebot could not see images. Bingbot could not see images. Alt text existed to tell those crawlers what the image was supposed to depict, and accessibility tools used the same field to describe images to users who could not see them.

That world ended in mid-2024 when GPT-4V shipped at scale, and the post-2024 architecture is fundamentally different. The current generation of multimodal AI models — GPT-4V, Claude 3 Opus Vision, Gemini 2.5 Pro Vision, and the Pinterest and Google Lens systems built on adjacent architectures — process the image and the surrounding text in a single attention pass. The pixels become a tensor. The alt text becomes a string token. The filename becomes another string token. The surrounding paragraph copy becomes additional context tokens. The model jointly attends across all of them and produces a semantic representation of what the image depicts and why it appears on the page.

This has three practical consequences.

First, pixel content and alt text are read together, not separately. If the alt text says "Glossier Cloud Paint blush in Puff" and the image shows a blush compact, the model unifies those signals into a high-confidence representation. If the alt text says "Glossier Cloud Paint" and the image shows a bottle of perfume, the model flags the mismatch and discounts both signals. Brands that historically wrote vague alt text — assuming the image would carry the meaning — now produce ambiguous signal that AI models discount.

Second, alt text resolves pixel ambiguity, which is more common than designers realize. A close-up of a beige liquid in a small glass bottle could be foundation, serum, hair oil, or olive oil. The pixel content is genuinely insufficient. The alt text becomes the disambiguator, and the brand that writes "Tata Harper Resurfacing Serum, 30ml glass bottle with dropper" gets cited correctly while the brand that writes "product" or "bottle" gets cited as something else or not at all.

Third, the surrounding DOM context still matters, but less than it used to. Pre-multimodal, AI systems had to infer image content from the surrounding text. Now they read the image directly and use the surrounding text as a confirmation signal. Brands that put their product images inside thin gallery components with no adjacent text used to lose ranking on that basis alone. In 2026, the loss is smaller but not zero — the entity context that surrounds the image still helps the model situate it in a category and a brand.

The next sections operationalize these dynamics into specific optimization patterns.

The Anatomy of an AEO-Ready Image

A product image optimized for visual AI search in 2026 has five coordinated surfaces. The brands that optimize all five compound the signal at every surface. The brands that optimize one or two lose most of the available lift.

The reason all five matter is that AI extraction pipelines read them sequentially and treat them as cross-validating signals. An image whose alt text, caption, filename, and schema all agree is treated as high-confidence. An image where the four surfaces disagree is treated as low-confidence and often dropped from the citation set entirely.

Take the canonical example of a Glossier Cloud Paint blush product image. The 2018 SEO version of this image had an alt text of "blush" or maybe "cloud paint blush" if the brand was thorough. The 2026 AEO version looks materially different.

The alt attribute reads: Glossier Cloud Paint cream blush in Puff, soft pink shade, 0.33 fl oz squeeze tube held against a neutral background. The caption reads: Cloud Paint in Puff is the brand's best-selling cool-toned pink, designed for buildable wash-of-color application on fair to medium skin. The filename reads: glossier-cloud-paint-puff-cream-blush-pink-0-33oz.jpg. The URL path lives at /cdn/products/cloud-paint/puff/. The schema.org ImageObject node includes contentUrl, caption matching the visible figcaption, creator set to the Glossier brand entity, and representativeOfPage set to true.

All five surfaces reinforce the same semantic content: this is a Glossier Cloud Paint blush in the Puff shade. When GPT-4V is asked "what is that pink blush Glossier sells in the squeeze tube," the model has high-confidence signal across every extraction layer and cites the correct product with the correct shade name. When Pinterest Lens scores a visual match against a user's photo of a similar product, the metadata reinforces the visual match. When Google Lens performs a reverse search, the structured data makes the citation cleaner.

The brands shipping this in 2026 — Glossier, Sephora's house brands, Goop, Allbirds, Outdoor Voices, the ecommerce arm of Sephora itself — are pulling away from competitors who treat alt text as a checkbox.

Writing Alt Text That Multimodal AI Cites

Alt text engineering is a writing discipline first, an SEO tactic second. The patterns that work in 2026 are surprisingly different from the patterns that worked under accessibility-only or Google-only optimization. The accessibility-only school recommends short, functional alt text that screen reader users can hear without fatigue. The Google-only school recommends keyword-dense alt text that reinforces the page's target query. Both schools produce alt text that underperforms in 2026 visual AI search.

The framework that works is what we call the Brand-Product-Attribute-Context pattern, or BPAC. Each component does a specific job in the multimodal extraction pipeline.

Brand anchors the image to the entity the AI model already knows. Without the brand, the model has to infer it from context, and inference is unreliable. Glossier Cloud Paint is unambiguous. Cloud Paint alone is ambiguous — multiple brands have used that or similar names.

Product specifies the product line or SKU. Cream blush is distinct from powder blush is distinct from liquid blush. The product term is what the AI model uses to categorize the image at the product-type level.

Attribute captures the variant — color, size, scent, material, model. Puff is the specific shade; 0.33 fl oz is the size. Attributes are how AI shopping queries get answered — when a user asks for the pink one or the small size, the attribute is what matches.

Context provides the use case or scene. Held against a neutral background tells the model this is a product shot, not an in-use shot. Applied to fair skin would indicate an in-use shot. Context disambiguates the image type and informs the AI model about how the brand wants the product framed.

The BPAC alt text reads as a complete declarative sentence: "Glossier Cloud Paint cream blush in Puff, soft pink shade, 0.33 fl oz squeeze tube held against a neutral background." This is 18 words. It is longer than the accessibility guidance traditionally recommends, but it is well within screen reader fatigue thresholds and well within the AI model's optimal context window for image captions.

The alt text antipatterns to avoid in 2026:

Keyword stuffing. Glossier Cloud Paint blush pink Puff cream beauty cosmetics makeup. This is what 2010-era SEO recommended. AI models in 2026 discount this aggressively as spam. The keyword density is the signal, and the signal says "low-confidence content optimized for crawlers."

Brand-only. Glossier. Used by many luxury sites that prefer minimalism. Provides no extraction signal beyond the brand entity and cedes citation share to competitors who write more specific captions.

SKU-only. GLO-CP-PUFF-001. Used by sites that auto-generate alt text from product database fields. AI models cannot interpret SKU strings without a lookup table and treat them as garbage tokens.

Empty alt. alt="". Used by sites where image alt has not been entered into the product CMS. Functionally equivalent to invisible to AI extraction and an accessibility failure that triggers WCAG 1.1.1 violations per the W3C accessibility guidelines. Google's own image SEO best practices documentation calls out missing alt as one of the highest-impact image SEO failures.

Caption duplication. Copying the visible caption into the alt attribute verbatim. Loses the available signal in one surface and degrades accessibility — captions are often phrased for human readers in ways that do not work as alt text.

The BPAC pattern avoids all five antipatterns and reads as natural language. It is the pattern we recommend across every ecommerce engagement we run in 2026.

The Schema.org ImageObject Stack

HTML alt text is necessary but not sufficient. The 2026 AEO advantage compounds when alt text is paired with explicit schema.org/ImageObject markup that structures the same semantic content for AI extraction pipelines that prefer structured data over inline HTML.

The full JSON-LD schema stack for AEO implementation covers the broader schema strategy. For images specifically, the implementation looks like this in 2026.

Every product image referenced from a Product schema node should be either inline ImageObject markup or a referenced ImageObject node with the following fields populated:

contentUrl. The canonical URL of the image asset. Should be a stable, indexable URL — not a hashed CDN URL that rotates on every deploy.

caption. The text caption matching the visible figcaption element on the page. Should be substantively identical to the rendered text, not a different string.

embeddedTextCaption. Added to the spec in 2025, this field carries the alt text as a first-class structured property per the schema.org ImageObject specification. Some AI extraction pipelines prefer this field over scraping the alt attribute from the HTML.

representativeOfPage. Set to true for the primary product image. Tells extraction pipelines this is the hero image and weights it accordingly.

creator. A Person or Organization node identifying the photographer or brand. For ecommerce, this is typically the brand entity itself.

license. URL pointing to the image license. For brand-owned imagery, this can be a custom license URL on the brand's domain. For Creative Commons imagery, the canonical CC license URL.

acquireLicensePage. For licensable images, the URL where licensors can acquire rights. Increasingly relevant for AI training data licensing.

width and height. Pixel dimensions of the image. Helps AI extraction pipelines understand image quality and aspect ratio.

The presence of complete ImageObject markup increases AI citation rates by approximately 41% over identical pages with no ImageObject markup, based on our 2026 audit. The brands shipping this consistently — typically those that have invested in a schema-generation pipeline at the CMS level — get cited at 2x to 3x the rate of brands relying on alt text alone.

Filename and URL Strategy

Image filenames are the most under-optimized surface in 2026 image AEO. The average ecommerce site we audit has filenames like IMG_4827.jpg, DSC_2156.jpg, or 2026-04-product-shot-final-v3.jpg. These filenames carry zero semantic signal and forfeit one of the cleanest extraction layers available.

The filename optimization pattern is straightforward: rename images to brand-product-attribute, hyphenated, lowercase, with a sensible URL path.

A blush product image filename should look like: glossier-cloud-paint-puff-pink-blush-0-33oz.jpg.

A serum product image filename should look like: tata-harper-resurfacing-serum-30ml-glass-bottle.jpg.

A jacket product image filename should look like: patagonia-down-sweater-jacket-black-mens-medium.jpg.

The naming convention should be enforced at the CMS or DAM layer, not as an editorial discipline that depends on individual contributors remembering to do it. Sites that ship a filename-generation rule at the upload step see immediate citation rate improvements within four to eight weeks as crawlers re-index the catalog.

The URL path should reinforce the brand and category hierarchy. Images served from /cdn/products/cloud-paint/puff/ provide stronger entity context than images served from /img/12345/abc.jpg. Sites that use UUID-based CDN paths can still optimize by exposing semantic redirect URLs that crawlers and AI models follow.

The cumulative impact of filename and URL optimization is real. Across our 2026 dataset, products with descriptive filenames and semantic URL paths were 18% more likely to appear in Pinterest Lens and Google Lens visual matching results and 24% more likely to be cited correctly by name when GPT-4V was asked to identify a similar product from a user-uploaded image.

Caption Strategy: Why Caption and Alt Should Not Match

One of the most common 2024-era image optimization mistakes was copying the visible caption verbatim into the alt attribute. The thinking was reasonable — both fields describe the image, why not unify them? In practice, the duplication forfeits half the available signal because AI extraction pipelines read the two fields as distinct semantic surfaces and apply different extraction weights.

The 2026 pattern is deliberate non-identical duplication. The alt text states the literal subject and brand in declarative form. The caption adds editorial context, comparison framing, or use-case content that the alt text cannot carry without becoming awkward.

For the Glossier Cloud Paint example:

The alt text is: Glossier Cloud Paint cream blush in Puff, soft pink shade, 0.33 fl oz squeeze tube held against a neutral background.

The caption is: Cloud Paint in Puff is the brand's best-selling cool-toned pink, designed for buildable wash-of-color application on fair to medium skin. The cream-gel formula sets to a natural-looking matte finish and works on cheeks, lips, and eyelids.

The alt text answers "what is this image of." The caption answers "why does this product matter and how is it used." Both surfaces feed the AI extraction pipeline. The non-identical pairing roughly doubles the citation surface area per image compared to identical duplication.

For category pages and editorial content, the pattern extends further. The alt text describes the image literally. The caption adds editorial commentary. The surrounding article paragraph provides comparison context, use case discussion, or brand positioning. Each layer is read independently and contributes to the model's representation of the image and its context within the broader content.

The 90-Day Image AEO Playbook

For ecommerce teams shipping image AEO infrastructure in the next quarter, the prioritized rollout:

1. Audit current alt text coverage. Crawl the full product catalog and produce a coverage report — what percentage of product images have alt text, what percentage have non-empty alt text, what percentage have alt text that includes the brand and product name. This baseline is the foundation for everything else. Most brands discover their coverage is 50 to 70% of where they assumed it was.

2. Define the BPAC standard. Document the Brand-Product-Attribute-Context pattern for your brand. Provide writers with three to five worked examples per product category — beauty, apparel, electronics, food. Make it the official editorial standard for all new image uploads.

3. Backfill the catalog. Prioritize the top 200 products by traffic and rewrite alt text to the BPAC standard. The 80-20 rule applies aggressively — for most ecommerce sites, 200 products generate 80% of citation-bearing queries. Backfill those first, then expand to the long tail.

4. Rename image filenames. Enforce a filename-generation rule at the CMS or DAM upload step. Rename existing high-traffic images to brand-product-attribute filenames. Set up 301 redirects from old image URLs to new ones to preserve any external link equity.

5. Ship ImageObject schema markup. Update product templates to emit complete schema.org/ImageObject markup with contentUrl, caption, embeddedTextCaption, representativeOfPage, creator, license, width, and height. Validate with Google's Rich Results Test and the schema.org validator.

6. Implement caption-alt deliberate non-duplication. Train writers on the pattern. The alt text states the subject; the caption adds editorial framing. Audit the top 200 products for caption-alt duplication and rewrite the caption to add use case or comparison context.

7. Optimize for Pinterest and Google Lens. For brands where visual discovery matters, ensure product images meet Pinterest's 2:3 aspect ratio recommendation at 1000x1500 minimum, and implement Pinterest Rich Pin meta tags. For Google Lens, complete Product schema with ImageObject is the primary requirement.

8. Instrument citation tracking. Track image-driven AI citations specifically — not just text citations. Tools like Profound and SerpRecon now expose image citation rates as a distinct metric. Build a weekly dashboard tracking the percentage of product-shaped queries where your images are cited as the visual reference.

9. Coordinate with the heading and chunking strategy. Image AEO compounds when the surrounding content is also extraction-friendly. The heading structure and chunking framework for LLM retrieval covers how to structure the text content that surrounds your images for maximum citation lift.

10. Audit quarterly. Image AEO is not a one-time project. As products launch and the catalog grows, the BPAC standard needs to be enforced continuously. Run a quarterly coverage audit and remediate any drift before it accumulates.

This sequencing reflects the actual deployment patterns of the brands we have worked with in 2026. The full rollout takes a determined ecommerce team about 90 to 120 days end to end. The citation rate improvements typically begin appearing in week six and compound through quarter two.

Vertical-Specific Patterns

The general BPAC pattern works across every ecommerce vertical, but several verticals have specific optimization requirements that meaningfully outperform the generic approach.

Beauty and cosmetics is the most image-dependent ecommerce vertical and the one where image AEO produces the largest citation lift. For a comprehensive treatment, see the beauty and cosmetics AEO playbook for AI product discovery. The vertical-specific additions are shade-name specificity, skin-tone context, and formulation type. A foundation alt text should include the brand, product line, shade name, undertone, and SPF rating. A skincare alt text should include the active ingredient, formulation type (serum, cream, oil, balm), and packaging format.

Apparel requires color, fit, material, and model context. A jacket alt text should specify the cut (slim, regular, oversized), the material (down, wool, technical shell), the color, and the model (men's, women's, unisex). The dual-language pattern matters for global brands — alt text in the primary market language plus localized versions for major secondary markets.

Electronics benefits from specification-bearing alt text. A laptop alt text should include the model name, screen size, processor generation, color or finish, and the configuration tier. Generic alt text like "MacBook Pro" loses to specific alt text like "Apple MacBook Pro 16-inch M4 Max in Space Black, 48GB memory, 1TB storage."

Furniture and home goods benefit from dimensional and material specificity. A sofa alt text should include the brand, model name, color and material, configuration (sectional, loveseat, three-seater), and dimensional context (compact, oversized, modular).

Food and beverage benefits from flavor, format, and dietary attribute specificity. A protein bar alt text should include the brand, flavor, format, dietary attributes (vegan, gluten-free, keto), and pack size.

The pattern across all five verticals is the same: the attribute layer is what drives AI shopping citations, because shopping queries are typically attribute-bearing. The brands that include the attribute layer in their alt text systematically outperform brands that include only brand and product.

What Visual AI Search Actually Looks Like in 2026

To make the optimization patterns concrete, here is what happens when a user searches visually in 2026 across the four major surfaces.

A user uploads a photo of a friend's blush product to ChatGPT and asks "what brand is this." GPT-4V reads the image tensor and extracts pixel features. It then performs a similarity search across its visual entity database and surfaces candidate matches. For each candidate, it weighs the visual similarity against the textual metadata of the candidate product pages — alt text, filenames, schema markup. The candidate whose textual metadata most strongly matches the visual features gets cited. A Glossier Cloud Paint page with BPAC alt text wins this comparison against a generic competitor with empty alt text.

The same user asks Pinterest Lens to find similar products. Pinterest's developer documentation on visual search describes how Lens runs a visual similarity match against the Pin index and surfaces visually matching results. For each result, the structured metadata from the linked product page is fetched and used to populate the Pin overlay — price, availability, brand. Pins linked to PDPs with complete Product schema and ImageObject markup get accurate overlays. Pins linked to PDPs with incomplete schema get incomplete overlays or get dropped from the result set.

The same user asks Google Lens. Google Lens performs a visual match and returns results sorted by a combination of visual similarity and structured data quality. Sites with complete Product and ImageObject schema rank higher. Sites with empty alt text and missing schema rank lower or are excluded.

The same user asks Claude Vision to compare two blush products. Claude reads both images plus the alt text plus the surrounding context. It generates a side-by-side comparison that quotes the brand and product names directly. Brands whose alt text correctly identifies them get correctly named in the comparison. Brands whose alt text is generic get described as "the pink blush on the left" — a citation failure that hurts brand recall.

In all four cases, the brands with engineered alt text and complete schema markup are the brands that win the citation. The brands relying on 2018-era image SEO are functionally invisible to the visual AI surfaces that increasingly drive product discovery.

Common Mistakes Brands Make in 2026

A short list of patterns we see repeatedly across underperforming brands:

Treating alt text as an accessibility-only field. Many design and engineering teams still view alt text as a screen-reader concern owned by the accessibility team. The 2026 reality is that alt text is dual-purpose — accessibility plus AI extraction — and the optimization patterns for both readers overlap substantially but are not identical. Brands that staff alt text writing as an accessibility-only function produce alt text that is functional for screen readers but suboptimal for AI extraction.

Auto-generating alt text from product database fields. A common shortcut is to concatenate the product name and SKU into the alt attribute programmatically. This produces alt text like "Cloud Paint - GLO-CP-PUFF-001" which has the BPA components but misses the C and reads as machine-generated. AI models discount auto-generated alt text relative to alt text that reads as human-written.

Relying on AI-generated alt text without review. The 2025 wave of CMS plugins that auto-generated alt text using GPT-4V was directionally helpful but is not sufficient. AI-generated alt text is generic by default — it describes what the image depicts without including brand, product, or attribute context that the brand owns. Brands need to layer brand-aware editing on top of any AI alt text generation.

Inconsistent visual style. Brands whose product photography varies wildly across the catalog — different backgrounds, different lighting, different framings — confuse the visual matching algorithms in Pinterest Lens and Google Lens. The brands that win visual discovery have disciplined visual systems and consistent product photography across the catalog.

Ignoring image freshness. AI models weight recently published or recently updated content more highly. Images that have been on a site for five years without re-encoding or metadata refresh are discounted. Brands that periodically refresh their hero product images and update the surrounding metadata maintain higher citation rates than brands with stale catalogs.

Missing the embeddedTextCaption schema field. The schema.org spec was updated in 2025 to support embeddedTextCaption as a structured way to associate alt text with the ImageObject node. Most schema implementations have not been updated to include it. Adding it is a small change that compounds across the catalog.

Forgetting Open Graph image tags. Image AEO is not just about the canonical PDP — it is also about the previews that appear when product pages are shared on social, in messaging apps, and increasingly in AI assistants that fetch Open Graph metadata. Complete OG image tags with og:image, og:image:alt, og:image:width, and og:image:height should be on every PDP.

The pattern across all seven mistakes is the same: brands treat image AEO as a checkbox rather than a discipline. The brands winning in 2026 treat it as a discipline that requires editorial, engineering, and product coordination.

Takeaway: Image alt text engineering is one of the highest-leverage AEO investments available in 2026, and it remains chronically under-resourced relative to the citation lift it delivers. Multimodal AI models read alt text, pixel content, captions, filenames, and structured data in a single pass, and brands that coordinate all five surfaces get cited at 2.7 times the rate of brands relying on 2018-era image SEO. The BPAC pattern — Brand-Product-Attribute-Context — is the writing discipline that works. The schema.org ImageObject stack is the structured-data layer that compounds the signal. The filename and URL strategy is the cleanup that closes the loop. Brands that ship the full playbook in the next 90 days will compound their citation lead through 2027 as visual AI search continues to absorb product discovery from traditional ecommerce surfaces. The window to build the infrastructure is now.