Key takeaways
- Most brands audit whether they appear in AI answers. Almost nobody audits where their citations come from — and that concentration is the real risk.
- Citation share is volatile at the domain level: Reddit went from 6.11% of ChatGPT citations in May 2026 to 0.52% by mid-August. If Reddit was carrying your visibility, you lost it overnight.
- A Source Diversity Score measures concentration across three axes: domains, content types, and platforms. Low diversity means high fragility.
- Single-snapshot audits are unreliable. Citation counts dropped ~27% across all ChatGPT models after the GPT-5.3 rollout in March 2026 — a platform change, not a content problem.
- You can compute a diversity score yourself with a spreadsheet and ~35 prompts, or use monitoring tools to track it continuously.
Here's a scenario that played out for a lot of brands in August 2026. Their ChatGPT visibility had been solid for months. A chunk of it — maybe most of it — came from Reddit threads ranking in AI answers. Then, on August 14, Reddit's share of ChatGPT citations collapsed from roughly 3.8% to under 1% in a matter of days. For brands whose AI visibility was riding on Reddit, that wasn't a dip. It was a wipeout they didn't see coming, because they were tracking the wrong thing.
Most AI visibility audits ask one question: are we mentioned? They report a visibility score, maybe a share of voice against competitors, and call it a day. What they almost never ask is: where do our citations actually come from, and how concentrated are they? That second question is what this framework is for.
I'm calling it a Source Diversity Score. The idea is borrowed from portfolio theory, of all places: if all your holdings are in one stock, you're not diversified, you're exposed. Same with AI citations. If 70% of the sources citing your brand are one domain, one content format, or one platform, your visibility has a single point of failure — and in 2026, AI search engines have been demonstrating exactly how fast those failure points can break.
Why citation concentration is the risk nobody is measuring
Let's start with the data, because the volatility here is genuinely striking.
Domain-level concentration is real and it moves fast
Promptwatch's citation share data for June 2026 shows Reddit leading ChatGPT Search citations at 3.71% — more than six times the runner-up, Wikipedia, at 0.56%. That sounds dominant, but here's the more important part: a month earlier, Reddit's share was 6.11%. A roughly 40% decline in a single month. And then in August, it fell off a cliff entirely — Promptwatch's Reddit citation tracking shows reddit.com's share of ChatGPT Search citations dropping from a steady ~3.8% to 0.52% by August 14, 2026, an 86% relative decline that coincided with ChatGPT's shift to using the site: operator at scale on August 8.
Google's surfaces behaved differently. AI Overviews declined gradually (2.37% to 2.10%, about an 11% relative drop) and AI Mode fell about 30% over the same period. No cliff, but the same direction.
The lesson isn't "Reddit is dead" — it's that no single-domain reliance is ever safe, and different engines de-risk differently. If your audit framework doesn't capture that, you're flying blind.
There's a counter-narrative worth flagging here. Yext's analysis of 6.8M AI citations found that 86% came from brand-managed sources (first-party sites, listings, reviews) once location context and query intent were applied, with Reddit and forums at just 2%. That's a wildly different picture from aggregate prompt sampling. The discrepancy matters: your citation mix depends heavily on your vertical, your query set, and your methodology. A national SaaS brand and a local dental practice will have completely different "natural" diversity profiles. Your diversity score has to be computed against your prompts, not industry averages.
Platform-level citation counts are controlled by the engines, not you
Here's something that trips up even experienced teams. Around the GPT-5.3 rollout on March 4, 2026, Promptwatch recorded a citation drop across all ChatGPT models simultaneously — average citations per response fell from about 6.4 to 4.7–4.9, a ~27% drop, with no recovery a month later. GPT-5.3, GPT-5.4, and GPT-5-Mini were all affected. That's a platform-level decision, not a content-quality signal.
If your audit is a single snapshot, you can't distinguish "our citations fell 27% because OpenAI changed something" from "our citations fell 27% because our content got worse." Those require opposite responses. This is why continuous monitoring beats point-in-time audits — a principle a University of St. Gallen study (arXiv, April 2026) reinforced when it found AI visibility behaves as a distribution, not a point estimate, due to the stochastic nature of LLM retrieval.
The same study measured citation concentration using the Gini coefficient and found it averaged around 0.71 across the engines they tested — with Google AI Mode concentrating citations most heavily and Perplexity distributing them most evenly. High Gini, in this context, means a few domains hoard the citations. That's your concentration risk, quantified by people who do this for a living.
Content-type concentration is a second, hidden axis
Diversity isn't just about which domains cite you. It's about what kind of content gets cited. Promptwatch's July 2026 citation type data shows product pages becoming the single most-cited content type in ChatGPT Search at roughly 32.8% of daily citations — nearly double their ~18% share in March. Listicles, news articles, how-tos, and social posts split most of the rest.
If your brand's AI visibility comes entirely from, say, listicles on third-party review sites, you're concentrated on a content type whose citation share is actively being squeezed by product pages. You'd want to know that before your traffic tells you.
The Source Diversity Score framework
Here's the framework. It's deliberately simple enough to run in a spreadsheet, because the point is to build the habit, not to buy software.
Step 1: Build your prompt set
Pick 30–50 prompts that reflect how real users ask about your category. Mix query types: informational ("what is X"), comparative ("X vs Y"), transactional ("best X for Y"), and troubleshooting ("how to fix X"). Everything-PR's 35-prompt citation share audit is a good template for structuring this across six query types.
Run each prompt across ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews. Record every citation source. You now have your raw dataset: a list of (prompt, engine, cited domain, content type) tuples.
Step 2: Compute concentration on three axes
For each axis, calculate the share of your citations going to each bucket, then compute a concentration index. I recommend the Herfindahl-Hirschman Index (HHI), the standard market-concentration metric: sum the squared shares of each source. HHI ranges from near 0 (maximally diverse) to 1.0 (all citations from one source).
| Axis | What you measure | Example question |
|---|---|---|
| Domain | Share of citations by cited domain | Is 60% of our visibility coming from two review sites? |
| Content type | Share by format (product page, how-to, listicle, forum post, news) | Are we only cited in listicles? |
| Platform | Share by AI engine | Does 80% of our visibility live in ChatGPT? |
Using DOJ antitrust conventions as rough thresholds: HHI below 0.15 is unconcentrated (healthy), 0.15–0.25 is moderate concern, above 0.25 is highly concentrated (fragile). For context, one industry analysis of AI citation markets reported the HHI jumping 293% from December 2025 to February 2026 — from 0.026 to 0.104 — meaning even the market-wide citation landscape was concentrating fast. Your brand-specific HHI can be far worse than the market's if you're dependent on one or two sources.
Step 3: Score and interpret
Combine the three axes into a simple report card:
| Diversity axis | HHI | Risk level | What it means |
|---|---|---|---|
| Domain | 0.08 | Low | Citations spread across many sites |
| Domain | 0.31 | High | One or two domains carry you — single point of failure |
| Content type | 0.22 | Moderate | Heavy on one format; watch for format-level shifts |
| Platform | 0.42 | High | Nearly all visibility in one engine; an engine change is existential |
A brand with high domain diversity but high platform concentration is in a different situation than one with the reverse. The first is exposed to engine algorithm changes; the second is exposed to a single publisher changing its site or going paywalled. Name the risk explicitly.
Step 4: Re-run monthly and watch the deltas
The score matters less than its trend. A domain HHI creeping from 0.12 to 0.19 over three months tells you concentration is building before it becomes fragility. Set thresholds that trigger action: if any single domain exceeds 40% of your citations, or any single engine exceeds 70%, that's your alarm bell.
Common concentration traps in 2026
Based on the data patterns above, here are the traps I'd check for first:
The Reddit trap. Reddit's citation share has been the story of 2026 — huge in ChatGPT, then collapsing. If your brand invested heavily in Reddit presence in 2025–2026 expecting durable AI visibility, Promptwatch's Reddit citation data is required reading. And note the mechanics: in January 2026, ~71% of Reddit posts ChatGPT cited had fewer than 10 upvotes, and ~60% of cited posts were six months or older. Reddit citations reflect historical community activity, not recent campaigns — you can't quickly rebuild this channel if it collapses.
The Google self-preference trap. In Google AI Mode, google.com alone captured 7.31% of citations in June 2026 — more than YouTube and Reddit combined — and Google-owned properties took over 10% of the mix. If your AI Mode visibility depends on ranking in Google's own properties, understand that this is a closed ecosystem Google controls end to end.
The single-engine trap. ChatGPT cites roughly 5 sources per web-search-enabled response; Google AI Overviews cites about 10, and Perplexity is remarkably consistent at almost exactly 10. Different engines, different citation economies, different risks. A brand visible only in Perplexity has a very different risk profile than one visible only in ChatGPT.
The crawler dependency trap. This one is infrastructure-level. Meta-WebIndexer went from ~2% to nearly 38% of all tracked AI crawler requests between mid-July and August 9, 2026 — a 17x increase in under a month, as Meta builds an independent web index. If Meta's crawlers can't reach your site, you're invisible to the next major AI engine before it even launches. Check your server logs and robots.txt for Meta-WebIndexer, Meta-ExternalAgent, and FacebookBot now, not later.
How to run the audit: manual vs. tooling
You can do this manually. Run your 35 prompts across five engines, log citations in a sheet, compute HHI per axis, repeat monthly. It's tedious but entirely doable, and honestly, doing it manually once teaches you more about your citation profile than any dashboard will.
For ongoing monitoring, tools help — mostly by giving you historical trend data you can't get from a one-off audit. A few options depending on your situation:
| Tool | Starting price | Best for |
|---|---|---|
| Promptwatch | $95/mo | Full-stack: citations, crawler logs, content gaps, automated fixes |
| Otterly.AI | $29/mo | Budget-friendly prompt tracking across ~6 platforms |
| Profound | $99/mo | Enterprise-grade conversation analytics (Claude tracking requires enterprise tier) |
| Peec AI | ~$95/mo | Mid-market, strong multilingual support |
| Rankscale | €20/mo | Broadest engine coverage (17+) plus a 200+ checkpoint page audit |
Promptwatch is the one I'd point most teams toward for this specific use case, because the diversity question is fundamentally about why you're visible, and its crawler logs show exactly which pages AI systems read, what errors they hit, and which citations actually flow from those crawls. Its citation analytics break down by domain, content type, and platform — which maps directly onto the three axes of this framework — and its Reddit and YouTube citation tracking covers channels most competitors ignore. It also monitors real UI outputs rather than just API responses, which matters because user-facing answers can differ from what APIs return.
If you're just starting and budget is tight, Otterly.AI gets you prompt-level tracking cheaply. If you're an enterprise wanting conversation-level analysis of what users actually ask, Profound's 400M+ real user prompt dataset is genuinely differentiated, though multi-engine tracking beyond ChatGPT requires their $399/mo tier.
One caveat on this whole category: pricing changes frequently. Profound's own materials quoted Growth at ~$499/mo in June 2026 while comparison sites listed $399/mo. Verify against vendor pages before committing.
What to do with a bad score
If your audit reveals high concentration, the fixes differ by axis:
Domain concentration: diversify your earned-media footprint. If two review sites carry you, invest in being cited across a broader set — industry publications, comparison pages, documentation, forums beyond Reddit. This is digital PR work, and it's slow, which is exactly why you should start before you need it.
Content-type concentration: if you're only cited in listicles, build content types that are gaining citation share. Product pages doubled their citation share between March and July 2026 — if your product pages are thin, that's a concrete, high-ROI fix. Structured product data, clear specs, FAQ sections on product pages.
Platform concentration: if 80% of your visibility is in one engine, the fix is understanding what the other engines cite and building for those citation patterns. Perplexity distributes citations most evenly (per the St. Gallen research), making it a good test bench for whether your content is inherently citable versus engine-specific.
And for the crawler side: verify that all major AI crawlers — ChatGPTBot, ClaudeBot, PerplexityBot, Google-Extended, Meta-WebIndexer — can access your key pages. Claude's citation crawler grew 100x in four months to a 1.73% daily share of tracked crawler traffic by April 2026. Being blocked from a compounding channel is a silent, compounding loss.
The bottom line
The brands that got hurt by the August 2026 Reddit collapse weren't lazy — they were measuring visibility without measuring fragility. A Source Diversity Score closes that gap. It takes an afternoon to set up manually, it surfaces risks that visibility scores hide, and in a year where citation share can move 40% in a month and platform-level citation counts can drop 27% overnight, it might be the most cost-effective risk management you do this year.
Start with the manual audit. Compute your HHI on all three axes. If the numbers scare you, that's the framework working.