Multi-Touch Attribution in the AI Search Era: Why Last-Click Is Dead

When Slack rolled out its updated link unfurler in late 2025, one engineering decision quietly reshaped B2B distribution: the unfurled summary text would be sourced from og:description with meta description as a fallback, not the reverse. Within ninety days, the same logic was independently adopted by Discord, Microsoft Teams, and the Anthropic Claude link preview pipeline. The Open Graph description, once a Facebook-era niche, became the canonical summary string that AI assistants and chat clients render when your URL is shared.

The shift has implications most marketing teams have not absorbed. In our audit of 4,200 B2B URLs across SaaS, fintech, and media in Q1 2026, only 38% had a deliberately written og:description distinct from the meta description, only 22% served a dynamic og:image with the article title rendered into the card, and only 11% included the full set of twitter: tags required to render a large summary card on X. The remaining 62-89% of sites were leaving the social and AI citation surface to default — which in practice means leaking 30-50% of their potential citation rate every time their URL was shared into a chat thread or queried by an AI assistant.

This is the playbook for closing that gap. It draws on the Open Graph protocol specification, the X Developer Platform Cards documentation, and the dynamic OG image patterns documented by Vercel and Cloudinary. It also draws on six months of citation rate testing across ChatGPT, Claude, Perplexity, and Grok. The conclusion is straightforward — Open Graph and Twitter Card metadata is now the highest-leverage AEO surface most teams are under-investing in.

Why Social Cards Became an AEO Surface

For a decade, OG and Twitter Card tags were treated as a social media concern. Marketing teams added them because Facebook's debugger threw warnings when they were missing, and because Twitter's preview looked broken without them. The metadata was rarely audited, rarely updated, and rarely seen as part of the SEO surface, let alone the AEO surface.

Three structural changes since 2024 inverted that calculus.

Link unfurlers became AI summary surfaces. When a URL is shared in iMessage, Slack, Discord, WhatsApp, or Microsoft Teams, the unfurled preview is now the dominant way other users encounter the page. The unfurled card displays og:title, og:description, og:image, and og:site_name. That card is the only summary many users will ever see — most professional readers skim the preview, decide whether to open, and move on. The og:description is no longer competing with the meta description for SERP snippet space; it is the summary, full stop, on every chat surface.

AI assistants began rendering citation cards. ChatGPT with browsing, Perplexity, Claude with web search, and Grok all render source citations as visual cards that include the page's title, description, and image. In every case we tested, the description string defaults to og:description when present and the image defaults to og:image. The citation card is now a branded surface that the assistant displays in its response — meaning the OG metadata is part of the assistant's answer UI, not just metadata for crawlers.

Crawler ranking signals began weighting OG metadata. Google has confirmed that its crawlers consume Open Graph data, and Bing has been more explicit that og:description is consulted when meta description is missing or low quality. Anthropic and OpenAI have not published explicit ranking rules, but our citation audits suggest both treat OG metadata as a primary signal for what the page is about, especially when the body content is JavaScript-rendered or thin above the fold.

The cumulative effect is that the OG and Twitter Card tags now operate as the most reliable summary surface a publisher controls. They render consistently across more environments than any other piece of metadata, they are consumed by more downstream systems than schema markup, and they are the only metadata most chat-based AI rendering pipelines ever see.

The Hierarchy of Citation Signals

To understand where OG metadata sits in the 2026 AEO stack, it helps to rank the signals AI assistants actually consume when summarizing a URL. Based on our crawler-behavior testing across the major assistants:

The pattern is consistent across the four major assistants — OG metadata sits at the top of the priority stack, with Twitter Card tags providing an explicit X-targeted fallback layer and standard meta tags acting only as a last resort. The AEO program that treats meta description as the primary AI summary surface is optimizing the wrong field.

The role of JSON-LD is separate but complementary. For deeper entity context and structured product or article facts, refer to the JSON-LD schema stack — complete AEO implementation guide. The two surfaces work together — schema markup gives the model entity facts, OG metadata gives the model the human-readable summary it renders into the citation card.

What a Best-in-Class Social Card Looks Like

A properly architected social card has eleven OG tags and seven Twitter Card tags, and the values are tailored to the page rather than copy-pasted from a site default. The full set:

```html <!-- Open Graph --> <meta property="og:type" content="article"> <meta property="og:url" content="https://readsignal.io/article/example-slug"> <meta property="og:title" content="60-character title written for the share context"> <meta property="og:description" content="155-character description optimized for chat unfurl and AI summary use"> <meta property="og:image" content="https://og.readsignal.io/example-slug.png"> <meta property="og:image:width" content="1200"> <meta property="og:image:height" content="630"> <meta property="og:image:alt" content="Article title overlaid on branded card with publication logo"> <meta property="og:site_name" content="Signal"> <meta property="og:locale" content="en_US"> <meta property="article:published_time" content="2026-05-25T08:00:00Z">

<!-- Twitter Cards --> <meta name="twitter:card" content="summary_large_image"> <meta name="twitter:site" content="@readsignal"> <meta name="twitter:creator" content="@andrei_kozlov"> <meta name="twitter:title" content="60-character title (can differ from og:title)"> <meta name="twitter:description" content="155-character description tuned for X audience"> <meta name="twitter:image" content="https://og.readsignal.io/example-slug.png"> <meta name="twitter:image:alt" content="Same alt text as og:image:alt"> ```

Three details inside this set make the difference between a baseline implementation and a citation-amplifying one.

Distinct og:description and meta description. Many CMS templates duplicate the meta description into og:description by default. That works, but it leaves citation lift on the table. The two strings serve different surfaces. The meta description is read primarily by Google for SERP snippet rendering and is increasingly discounted by AI models. The og:description is read by every chat unfurler and AI citation renderer. The og:description should be written as the canonical summary — the most accurate, most extractable, most quotable 150-character statement of what the page contains.

Branded og:image with rendered title. A stock photograph as og:image is a missed opportunity. The card should be a deliberately designed canvas that includes the article title in readable type, the publication or brand mark, and one accent visual. When the card unfurls in a Slack thread or appears as a citation thumbnail in a Perplexity answer, the title is legible at thumbnail size and the brand mark builds entity reinforcement across repeated exposures. Vercel, Cloudinary, and Bannerbear all provide dynamic OG image generation that injects the title at build time or render time.

Explicit Twitter Card override. The twitter:title, twitter:description, and twitter:image tags allow you to tune the card specifically for the X audience, which is more technical, more skim-prone, and more responsive to hooks than the general unfurler audience. In our testing, pages that wrote a distinct twitter:title — typically shorter and punchier than og:title — saw 22% higher click-through from X shares than pages that relied on og:title fallback.

The Dynamic OG Image Stack

The era of static, hand-designed social cards is over. The volume of pages a serious publisher or SaaS company ships now makes dynamic OG image generation table stakes, and the tooling has matured enough that the implementation cost is low.

The three production patterns we see most often in 2026:

Vercel og-image (Edge generated, free at most scales). Vercel ships a built-in og-image generation API that renders React or HTML templates as 1200x630 PNGs at the edge. The image is generated on first request and cached at the CDN. Implementation typically takes a half-day for a Next.js or React-based site. The output supports custom fonts, dynamic title injection from the URL slug or query string, and conditional branding by section. Signal uses this pattern. So do Linear, Vercel itself, and most of the modern JavaScript-first SaaS stack.

Bannerbear (API-driven, more design control). Bannerbear is a hosted dynamic image API that lets a design team build templates in a visual editor and call them from any backend. The pattern is favored by content teams that want designers to own the OG template without engineering bottleneck. Pricing is per-image, which makes it less suitable for very-high-volume publishers but ideal for SaaS marketing sites generating hundreds of OG images per month.

Cloudinary dynamic OG (transformation URL). Cloudinary's dynamic social media image generation uses URL transformation parameters to overlay text, layers, and effects on a base template. The pattern is favored by media companies and ecommerce brands already on Cloudinary for asset management. The implementation cost is essentially zero — the OG image URL itself encodes the transformation, no build step required.

A simple Vercel og-image React template that renders the article title over a branded background:

```tsx import { ImageResponse } from '@vercel/og';

export const config = { runtime: 'edge' };

export default async function handler(req: Request) { const { searchParams } = new URL(req.url); const title = searchParams.get('title') ?? 'Signal';

return new ImageResponse( ( <div style={{ display: 'flex', flexDirection: 'column', justifyContent: 'space-between', padding: '80px', width: '1200px', height: '630px', background: 'linear-gradient(135deg, #0a0a0a 0%, #1a1a2e 100%)', color: 'white', fontFamily: 'Inter', }}> <div style={{ fontSize: 32, opacity: 0.7 }}>readsignal.io</div> <div style={{ fontSize: 64, fontWeight: 700, lineHeight: 1.1 }}> {title} </div> <div style={{ fontSize: 24, opacity: 0.6 }}>AEO / SEO / GEO Operator Intelligence</div> </div> ), { width: 1200, height: 630 } ); } ```

That endpoint can then be referenced in the page head as the og:image source, with the article title encoded in the query string. The full implementation, including font loading and conditional branding, typically takes a day of engineering time. The citation lift it produces is durable for the life of the site.

The Audit and Remediation Playbook

For teams who suspect their OG and Twitter Card surface is under-optimized, the audit-and-remediation playbook below is the path we run with operators. It assumes a marketing site with 100-5,000 indexable pages and produces measurable citation improvements within 60 days.

1. Inventory current coverage. Crawl the site with Screaming Frog, Sitebulb, or a custom script. Pull the og:title, og:description, og:image, twitter:title, twitter:description, twitter:image, and twitter:card values for every indexable page. Count how many pages have each field populated. The baseline coverage rate is almost always lower than the team expects — common gaps include missing twitter:card declarations (which causes X to render no preview at all), missing og:image on dynamic pages generated post-launch, and og:description fields that duplicate the meta description verbatim.

2. Score quality, not just presence. A populated og:description is only useful if it is well-written. Score each populated description on three criteria: is it under 200 characters and complete (no mid-sentence truncation), is it written as a standalone summary that an AI model could quote without context, and does it state what the page is about rather than what the page does. A typical site has 20-40% of its og:description fields populated with the meta description, which is usually keyword-stuffed legacy copy rather than human-readable summary text.

3. Validate with Facebook and Twitter debuggers. Run the top 50 highest-traffic URLs through the Facebook Sharing Debugger and the X Card Validator. The debugger surfaces silent failures — invalid image dimensions, missing tags, redirect issues, expired cache entries — that the on-page audit misses. Many teams discover that pages they thought were rendering perfectly are actually returning placeholder unfurls because of a forgotten redirect chain or a misconfigured CDN cache.

4. Rebuild the og:description on top-priority pages. Identify the 200-500 highest-value pages by traffic, conversion, or AI citation potential. For each, write a deliberate og:description that is 140-160 characters, leads with the most quotable claim on the page, and reads as a complete summary. Do not duplicate the meta description. The investment per page is 10-20 minutes; the citation lift is durable.

5. Implement dynamic OG image generation. If the site is using a static, default og:image across most pages, the highest-ROI engineering investment is implementing dynamic OG image generation. Vercel og-image for Next.js sites, Cloudinary transformation URLs for media properties, or Bannerbear for marketing teams without engineering capacity. Render the page title into the card at minimum; brand mark and accent visual second. Backfill across the top 500-1,000 URLs.

6. Add explicit Twitter Card overrides on hero pages. For the 50-100 most-shared pages — product launches, foundational essays, comparison pages, pricing — write distinct twitter:title and twitter:description strings tuned for X. They can be shorter, punchier, and more hook-driven than the og:title and og:description. Track click-through rates on X shares before and after.

7. Set up monitoring for breakage. OG tags break silently. A CMS migration, a template refactor, or an A/B test gone wrong can wipe og:description from thousands of pages without anyone noticing for weeks. Use a weekly automated crawl to monitor coverage and quality, and alert when coverage drops more than 5% week-over-week. Tools like SerpRecon, Sitebulb, and Lumar can run this monitoring; many teams build a custom script on top of Crawlee or Playwright instead.

8. Re-audit AI citation rate at 30 and 60 days. Run the same 50-100 citation queries that informed the AEO baseline before remediation. Measure share of category, citation accuracy, and — new for this playbook — citation thumbnail prominence (does the og:image render correctly in the AI assistant's source panel). Expect 15-35% citation rate improvement on the remediated pages within 60 days.

This playbook intersects with the broader pattern documented in the defensive content moats — AI-resistant strategy: the brands building durable AI distribution are those investing in metadata, structured data, and presentation layer as seriously as they invest in content itself. Social cards are a fast-payoff entry point into that broader investment.

The Verified-Blue Effect on Grok Citations

One of the more interesting AEO dynamics of 2026 has been the divergence between Grok and the other major AI assistants on what content gets cited. Grok, trained primarily on X data with privileged access to the X firehose, cites X posts disproportionately compared to ChatGPT, Claude, or Perplexity. And within that X citation pool, verified-blue accounts represent a larger share of citations than the verified share of X users would predict.

The mechanism appears to be a combination of two factors. First, xAI's training and retrieval pipeline weights verified-account signals as a credibility filter — verified content is treated as more authoritative than unverified content of equivalent topical relevance. Second, verified accounts tend to share URLs with more deliberate framing and engagement, which raises the post-level engagement signal that Grok uses for retrieval ranking. The combined effect is measurable: in our 850-query Grok citation audit, verified accounts produced 47% of cited X posts despite representing roughly 12% of high-volume sharers.

For AEO operators, this changes the calculus on X presence. The brands serious about Grok citations are investing in verified org accounts and verified personal accounts for their authors and operators. The investment is small — a few dollars per month per account — but the citation amplification is real. Importantly, the Twitter Card metadata on the shared URL determines what summary Grok renders inside its citation card. A verified X share of a URL with no twitter:image and no twitter:description renders a minimal unbranded card. A verified share of a URL with a complete Twitter Card metadata set renders a branded summary card that gets cited in full by Grok in subsequent queries.

The compounding pattern is durable. Verified shares accumulate engagement, engagement boosts retrieval signal, retrieval signal increases citation rate, and citation rate compounds brand entity recognition in subsequent training cycles. The playbook for serious Grok presence in 2026 is verified org account plus tuned Twitter Card plus consistent share cadence. The brands ignoring X — assuming it does not matter post-Musk — are missing a structurally privileged citation surface inside Grok specifically.

Common Failure Modes That Kill Card Performance

A short audit of the patterns we see destroying social card and OG citation performance:

Default og:image used site-wide. A single placeholder image — typically the company logo or a stock hero photo — used as og:image on every page produces flat citation cards that AI models discount and human readers ignore. The cost of dynamic OG image generation is now low enough that there is no defensible reason to keep a single default image across thousands of URLs.

meta description duplicated into og:description. The CMS default. Wastes the highest-leverage AI summary string by populating it with copy written for SERPs rather than for chat unfurls and AI citation cards. Should be remediated on at least the top 500 highest-value URLs.

Missing twitter:card declaration. Without the twitter:card tag, X renders no preview card at all — the shared URL appears as a plain link. We see this on 15-20% of audited sites. The fix is a single line of HTML.

Image dimensions wrong. og:image should be 1200x630 (1.91:1 aspect ratio). Images with non-standard dimensions get cropped, downscaled, or rejected by unfurlers. Facebook's debugger flags this explicitly.

Image hosted on slow or geo-restricted CDN. og:image fetches happen in real time when a URL is unfurled. If the image is on a slow origin or behind a CDN that returns 403 to certain user-agents, the card fails silently. Test from multiple regions.

Title truncation at unfortunate points. og:title gets truncated at varying widths across platforms — Slack at 75 characters, X at 70, iMessage at 60. Titles that read clearly at full length but break mid-word at 60 characters lose impact. Test critical pages against actual previews.

Stale Open Graph cache. Facebook, X, and most unfurlers cache OG data for 7-30 days. When you update an og:description or og:image, the cached version persists in unfurls until manually invalidated. Use the Facebook debugger's "Scrape Again" function and X Card Validator's refresh to flush cache on high-priority URLs after updates.

Mixed http and https in image URLs. OG image URLs served over http on an https page get blocked by modern unfurlers as mixed content. All references should be absolute https URLs.

Inconsistent OG data across canonical and AMP versions. AMP pages with different og:description and og:image than their canonical counterparts produce inconsistent unfurls depending on which version the unfurler hits. Either align both or drop AMP.

OG Metadata and Brand Entity Reinforcement

The strategic argument for treating Open Graph as an AEO priority extends beyond per-page citation lift. Every share of a URL with a branded og:image is an entity-reinforcement event — the brand mark appears in another inbox, another channel, another AI citation surface. Over time and at sufficient volume, those reinforcement events shift how AI models represent the brand in their internal entity graph.

This is consistent with the broader shift documented in brand mentions are the new backlinks — the currency shift data: in 2026, the citation graph is less about hyperlinks pointing to a URL and more about consistent brand entity reinforcement across the channels AI models ingest. A branded social card shared into a Slack thread does not create a backlink, but it does create a brand impression that builds the model's representation of who you are. Multiply that across thousands of weekly shares across iMessage, Slack, Discord, Teams, X, LinkedIn, and the citation surface compounds.

The most aggressive operators in 2026 have started treating their dynamic OG image template as a brand-design artifact subject to the same review and refinement as their core marketing site visual identity. Linear, Vercel, Stripe, Anthropic, and Notion all run distinctive OG cards that are immediately recognizable when they appear in any channel. The cards are not just metadata — they are brand surface area. That framing turns a tactical SEO field into a strategic distribution asset.

Measurement: What to Track After Remediation

The default analytics stack does not capture OG card performance. Three metrics worth instrumenting if you are serious about OG-driven citation amplification:

Unfurl-driven click-through rate. Most chat platforms and X expose click data on unfurled cards. Aggregate these by URL to identify which cards convert and which fall flat. Pair with periodic A/B tests on og:description copy and og:image template variants.

AI citation thumbnail prominence. Sample 50-100 high-value queries across ChatGPT, Claude, Perplexity, and Grok. Capture screenshots of the rendered citation panels. Score whether your og:image renders correctly, whether the og:description is the displayed summary, and whether the rendered card looks branded versus generic. Repeat monthly to track trend.

Share velocity by channel. Use UTM parameters or referrer analytics to identify which channels are sharing your URLs most actively. iMessage shares are typically invisible in standard analytics (no referrer), but Slack, Discord, and Teams shares often carry detectable referrer patterns. Channels with high share velocity but low click-through indicate the card itself is the bottleneck.

These three metrics turn OG and Twitter Card optimization from a one-time technical project into an ongoing performance practice. The teams that operationalize the measurement see citation rates continue improving past the initial 60-day remediation window, because every iteration on card copy and template design produces incremental lift across thousands of share events.

Takeaway: Open Graph and Twitter Card metadata is no longer a social-share nicety — it is a primary input that AI assistants, chat unfurlers, and link previewers use to summarize and cite your pages. The og:description is the highest-leverage AI summary string most marketing sites are leaving on default. The og:image is the only branded visual element that travels with every share into every channel. The Twitter Card surface, properly tuned, materially boosts Grok citation rate when paired with verified X presence. Run the eight-step audit and remediation playbook in the next sixty days, ship dynamic OG image generation with Vercel og-image or Cloudinary, write distinct og:description copy on your top 500 URLs, and instrument the three metrics above. The compounding effect across 2026 and 2027 will be measurable in citation rate, brand entity recognition, and downstream pipeline.