AI Keeps Citing Your Business. What It Says About You Changes More Often.
Most AEO work is structured around a single question: does AI mention your business? The fix plans, the audit scores, the monitoring dashboards -- they measure whether your name appears in a response. That is a useful metric. It is not the whole picture.
In our November 17, 2026 research filing (Scout session 202), we documented findings from arXiv:2606.20065 -- Pratyush Kumar's large-scale GEO study covering 100,000+ AI prompt responses across ChatGPT, Claude, Perplexity, and Gemini. The finding relevant to anyone running AEO audits: sentiment framing changes 6.7x more often than whether a business is cited at all.
A business can have a stable citation presence and completely inconsistent framing at the same time. These are two different axes, and fixing one does not fix the other.
What "Sentiment Framing" Actually Means
Citation presence is binary: either your business appears in an AI response or it does not. Sentiment framing is what the AI says once it mentions you -- whether the description positions you as a credible recommendation, a neutral listing, or a business with notable concerns.
For a plumber, citation presence is "Murphy Plumbing in [city]." Framing is what follows: "well-reviewed, available for emergency service" is positive framing. "Pricing not widely documented" or "mixed customer reviews" is negative framing. Both responses cited the business. Only one is likely to generate a call.
The 6.7x figure means the words around your citation change roughly seven times as often as whether you appear at all. Infrastructure fixes -- directories, schema, NAP consistency -- move your citation presence. They do not directly move how that citation is worded.
Why Citation Presence Is Sticky and Framing Is Not
Citation presence is driven primarily by entity infrastructure: how well-established your business is across the sources AI platforms retrieve from. This includes directory presence, review platform coverage, brand stature, and how clearly structured your entity record is across the web.
The arXiv:2606.20065 study documents the brand stature tiers: global brands are cited in roughly 73% of relevant AI queries. Mid-market brands with moderate third-party presence land at 44%. Local and niche businesses -- most of Sourcepull's clients -- start at approximately 11%. These percentages are relatively stable over short periods. Getting a business from 11% to 44% is Phase 1 and Phase 2 infrastructure work. The number moves slowly in response to real changes.
Sentiment framing operates differently. It is sensitive to whatever sources AI retrieves for a specific query at a specific moment. If a new negative review appears on a high-authority platform, it enters the retrieval pool and can shift how AI describes the business, even if nothing else changed. The same retrieval variability that makes citation drift a problem (a business cited today may not be cited next month for the same query) also creates framing variability: a business described one way today may be described differently tomorrow, without any intentional change on the business's part.
The 6.7x ratio reflects this structural asymmetry: infrastructure (the citation presence driver) is durable. Source pool content (the framing driver) is not.
The Content Alignment Factor
In our November 17, 2026 filing of the Discoveredlabs statistical analysis (Scout session 202, 2M citations across 10,000 pages, six months of data), the strongest predictor of citation quality was prompt-content alignment -- how directly the page content answered the specific query being asked.
This finding matters for framing because alignment does not just determine whether a page gets cited. It determines what gets extracted. A service page that speaks in general professional language ("comprehensive HVAC solutions for residential and commercial clients") matches weakly against buyer queries like "what does furnace installation cost in [city]" or "how fast can HVAC repair happen on a Sunday." When a weakly-aligned page is cited, AI retrieves whatever it can extract -- often a vague description of services that doesn't answer what the buyer asked. That is neutral-to-negative framing regardless of the business's actual quality.
A page written in buyer incident language -- "furnace installation costs in [city] typically run $3,500-$6,500 depending on unit size and ductwork condition" -- gives AI a specific, extractable answer. When that page is cited, the response reflects the content. The business gets described as a source that can answer the question, which is positive framing by definition.
The content alignment finding explains why two businesses with identical citation presence scores can get described completely differently. The score measures whether they appear. The description reflects what the pages actually say.
Review Quality and the Framing Floor
There is a point below which framing stops being volatile and becomes fixed negative: the star rating floor. Our research on ChatGPT citation eligibility for home services businesses documents a 4.0-star floor below which businesses are largely filtered out of recommendation responses. At 4.1 stars, citation presence is restored but framing tends to be generic or qualified -- AI platforms are reflecting what the third-party review record says, and a 4.1 average says "acceptable but not standout."
Our home services citation data (knowledge/home-services-ai-citation-data.md) documents a higher bar for competitive markets: businesses cited in recommendation contexts in competitive metros average 4.3+ stars, and businesses with strong competitive positioning average 4.6+. The gap between "not excluded" (4.0) and "described as a credible recommendation" (4.6+) is the framing range where most businesses operate.
Sentiment volatility below 4.6 stars is real and documented: a business sitting at 4.2 stars can get positive framing when a retrieval pulls its best-reviewed platform coverage and negative framing when it pulls a lower-scoring source, all within the same week. Infrastructure work (more reviews, platform consistency) compresses that variance, but it does not eliminate it until review quality consistently crosses the 4.3-4.6 threshold.
Two Separate Audit Dimensions
The practical implication is that any AI visibility audit has two distinct output dimensions, not one.
The first is citation presence: how often does the business appear in relevant AI responses? This is what most audit scores measure. It is a function of entity infrastructure and brand stature. It moves slowly in response to Phase 1 and Phase 2 fixes.
The second is sentiment framing: when it appears, how does AI describe it? This is a function of content alignment, review quality, and what the retrieval pool contains. It moves faster -- both in response to positive improvements and in response to new negative content entering the ecosystem.
A business that scores 7/10 on citation presence and gets described negatively in 40% of its citations has a different problem than a business that scores 3/10 on citation presence and gets described well when it appears. The fix plans are different. The timelines are different. The monitoring cadence is different.
The 6.7x ratio from arXiv:2606.20065 is a calibration tool for client conversations. When a business owner asks why their AI visibility score improved but they're still getting unfavorable descriptions from ChatGPT, the answer is not that the fix plan failed. It is that citation presence and sentiment framing are different axes, and they had not yet addressed the second one.
What You Can Fix
Citation presence improves through Phase 1 infrastructure: directory listings, NAP consistency, schema markup, review platform coverage. These signals accumulate over months.
Framing quality improves through content alignment and reputation signal strength. For Phase 2, that means rewriting service pages in buyer query language, adding pricing context for cost-category queries, ensuring FAQ content answers the specific questions buyers bring to AI, and maintaining a review quality floor that the retrieval pool reflects positively. For review quality specifically, the work is continuous -- new reviews change what the retrieval pool says about the business, which is the fastest-moving input to framing.
If you have run a Signal Check, the response text is not just a citation presence indicator -- it is a sample of current framing. Reading what the AI actually says about the business, not just whether the name appeared, is the diagnostic step that most audits skip.
See how your business scores on AI platforms.
Check your score — free