---
title: "How to Prioritize an AI Visibility Budget: What 110,877 Citations Say About Where the Money Should Go"
description: "CMOs now put 15.3% of marketing budgets into AI. The Machine Relations Index shows where AI citations actually land across nine source classes, why the mix inverts by category, and how to allocate against observed evidence instead of averages."
canonical: https://authoritytech.io/blog/how-to-prioritize-ai-visibility-budget-2026
last-updated: 2026-09-18
---

# How to Prioritize an AI Visibility Budget: What 110,877 Citations Say About Where the Money Should Go

CMOs now put 15.3% of marketing budgets into AI. The Machine Relations Index shows where AI citations actually land across nine source classes, why the mix inverts by category, and how to allocate against observed evidence instead of averages.

Canonical URL: https://authoritytech.io/blog/how-to-prioritize-ai-visibility-budget-2026
Published: 2026-09-18
Author: Jaxon Parrott
Topic: Machine Relations

Prioritize an AI visibility budget by the source classes AI engines actually cite in your category, not by channel habit. Today's [Machine Relations Index](https://machinerelations.ai/index) release attributes 110,877 cited answer runs across 22,179 domains to nine source classes. Editorial publications take 12.0%. Vendor-owned sources take 10.05%. Wire and press-release distribution takes 0.20%. Your category almost certainly does not match that average, and that is the first thing to check.

I want to be exact about what this is and is not. It is an observed map of where citations land. It is not a return-on-spend curve, and nobody has one.

## The number that should set the budget

Gartner's 2026 CMO Spend Survey put 401 marketing leaders on record: marketing budgets sit at [7.8% of company revenue](https://markets.ft.com/data/announce/detail?dockey=600-202605111155BIZWIRE_USPRX____20260511_BW321750-1), barely moved from 7.7% in 2025, and an average 15.3% of those budgets now goes to AI initiatives. Only 30% of those organizations report the maturity to scale what they are funding. Inside the same survey, [martech fell to a five-year low of 19.4%](https://www.chiefmarketer.com/gartner-cmo-spend-survey-budgets-reflect-increase-in-consumption-based-martech-paid-media-spend/) of budget while paid media hit a five-year high of 31.4%.

So the allocation question is already live in most companies. What is missing is a denominator. A team deciding where AI visibility money goes is usually choosing between "more content", "more PR", "more tooling" and "more paid", with no observed evidence about which source classes the engines actually draw from.

That evidence exists now, and it is public.

## What the index actually measures

Release `mri_score_v2.0+2026-09-18+8fa38e54dd0a`, artifact SHA-256 `8fa38e54dd0af0c5486b431b9ac3a8ca2c64159700c43094d2243cc005dfb988`, observation window 2026-05-10 through 2026-09-18, 125 days observed. Across that window the index recorded 124,397 source events over 15,782 answer runs, across six engines, and classified 22,179 cited domains into nine source roles.

The role table below counts cited runs: how many observed answer runs cited at least one domain in that class. The right-hand column is the one most people have never seen.

| Source class | Domains | Cited runs | Share of cited runs | Cited runs per domain |
|---|---:|---:|---:|---:|
| Other observed source | 18,623 | 67,043 | 60.46% | 3.6 |
| Editorial publication | 1,222 | 13,309 | 12.00% | 10.8 |
| Vendor-owned source | 823 | 11,153 | 10.05% | 13.5 |
| Market and company database | 688 | 6,066 | 5.47% | 8.8 |
| Community and social platform | 27 | 4,314 | 3.89% | 159.7 |
| Academic and government source | 413 | 3,872 | 3.49% | 9.3 |
| Analyst and consulting research | 364 | 3,381 | 3.04% | 9.2 |
| Search or media platform | 10 | 1,513 | 1.36% | 151.3 |
| Wire and press-release distribution | 9 | 226 | 0.20% | 25.1 |

Read the first row before any other. Three in five cited runs land on a source the classifier has not yet placed in a named class. That is the honest state of the map, and it caps how far any share number below it can be pushed. A team that treats "editorial is 12%" as "editorial is 12% of the opportunity" has misread the table. The named classes cover 39.5% of cited runs between them. Everything here describes that share.

## Concentration is the finding, not share

Share tells you where citations land in bulk. Cited runs per domain tells you what a single domain in that class is worth, and the spread is enormous.

Community and social platforms hold 27 domains and 4,314 cited runs: 159.7 cited runs per domain. Search and media platforms hold 10 domains at 151.3. Editorial publications hold 1,222 domains at 10.8. The long unclassified tail sits at 3.6.

A community or platform domain is cited roughly fifteen times as often as the average editorial domain and forty-four times as often as the average unclassified one. That is not a recommendation to buy community presence, and it cannot be bought anyway. It is a structural fact with a direct budget consequence: in the classes where a handful of domains absorb enormous citation volume, there is no long tail to work. You are either present on those few surfaces or you are not, and spreading budget thinly across many properties in those classes buys nothing.

The inverse holds for editorial. 1,222 editorial domains at 10.8 cited runs each is a real long tail, which means placement selection matters more than placement volume. Twelve percent of cited runs spread across 1,222 domains is a market where the specific outlet is the whole decision.

## Wire distribution is 0.20% of the map

Nine wire and press-release domains. 226 cited runs. Two tenths of one percent of role-attributed citations.

I have argued against the wire for years on the grounds that it was never earned attention, so treat this as a finding I am motivated to like and check it yourself against the release. The number is what it is: in this window, in this index, press-release distribution is the least-cited named source class, below academic sources, below analyst research, below market databases.

There is a real caveat and it cuts both ways. Wire domains score 25.1 cited runs per domain, above editorial's 10.8, because there are only nine of them. Concentration is high and total contribution is tiny. A wire release can still do the jobs wire releases do, including disclosure, syndication pickup and creating a datable public record. What the Machine Relations Index does not establish is a case for funding the wire as an AI citation strategy.

## Your category inverts the average

This is the part that makes portfolio averages dangerous. The role mix below is computed across the 100 most-cited domains in each category, so its denominator is that top 100, not the whole index.

| Category | Leading class | Lead share | Editorial | Vendor-owned | Community | Wire |
|---|---|---:|---:|---:|---:|---:|
| Cybersecurity | Vendor-owned | 32.3% | 21.0% | 32.3% | 8.2% | 0% |
| Enterprise software | Vendor-owned | 34.9% | 14.1% | 34.9% | 7.5% | 0% |
| Healthcare services | Editorial | 26.5% | 26.5% | 11.6% | 6.7% | 0.5% |
| HR and talent | Editorial | 22.2% | 22.2% | 15.8% | 8.0% | 0.7% |
| Consumer finance | Unclassified | 57.7% | 23.7% | 3.6% | 7.3% | 0% |
| Fintech | Unclassified | 34.7% | 14.4% | 24.7% | 6.2% | 1.0% |
| Deep tech and hardware | Unclassified | 63.8% | 5.9% | 1.0% | 12.9% | 0% |
| AI visibility and GEO | Unclassified | 62.9% | 6.6% | 10.7% | 10.5% | 0% |

Four readings follow, and each one changes an allocation.

**The leading class flips.** Vendor-owned documentation leads cybersecurity and enterprise software. Editorial publications lead healthcare services and HR. A single "invest in earned media" or "invest in your docs" instruction is wrong in half of these markets.

**Vendor-owned is not uniform.** It runs from 34.9% in enterprise software to 1.0% in deep tech and hardware and 0% in the top 100 for consumer health and consumer products. In enterprise software, your own well-structured documentation is a primary citation asset. In deep tech, on the same evidence, it is close to invisible.

**Editorial share has a four-fold range.** 26.5% in healthcare services against 5.9% in deep tech and hardware. The same earned media program is a different investment in each.

**The unclassified share is itself a signal.** Categories where it runs above 60% are markets whose cited sources do not resolve into recognizable institutional classes. AI visibility and GEO is one of them, at 62.9%, which is worth sitting with if you sell into that market.

## What the outside research agrees on

Three independent findings line up with the concentration reading, and none of them was produced by anyone selling a visibility service.

A [37,000-run audit across 215 commercial prompts and 19 sectors](https://arxiv.org/abs/2605.27439), evaluated against a 533-brand catalog in five prominence tiers, found the failure mode differs sharply by tier. Category leaders appear in nearly every relevant retrieval but win only 25 to 41% of the recommendation slots they reach. Challengers convert best at 37 to 52%. Mid-market brands drop to 88% coverage and 34 to 40% conversion. Specialists and regional players never surface at all in 48 to 52% of runs. The authors' conclusion is the budget conclusion: "No uniform optimization recipe wins; the right marketing investment depends on where the brand sits on the prominence ladder."

The [Discovery Gap study](https://arxiv.org/abs/2601.00912) tested 112 Product Hunt startups across 2,240 queries against ChatGPT and Perplexity. Recognition by name reached 99.4% and 94.3%. Discovery by category question collapsed to 3.32% and 8.29%. Optimizing site content for AI showed no correlation with discovery. What correlated with Perplexity visibility were referring domains (r = +0.319, p < 0.001) and community presence (r = +0.395, p = 0.002). I will not overstate it: that paper's own recommendation is to build the search foundation first and let model visibility follow, not to buy placements. Its useful contribution here is the direction of the signal, which is external corroboration rather than owned output.

Retrieval sets also differ from search results. A comparison of Google Search, Gemini and AI Overviews found [less than 0.2 Jaccard similarity](https://arxiv.org/abs/2604.27790) between what traditional search retrieves and what generative engines retrieve, and a separate study found [30% of AI Overview sources do not appear in first-page results](https://arxiv.org/abs/2605.14021). Budget inherited from an SEO plan is budget aimed at a different source set.

## The allocation method

Five steps. None requires a vendor.

**One. Find your category's mix, not the average.** Open your category on the index and read the source-role composition of its top cited domains. The portfolio table above is a starting point only if your category is not covered.

**Two. Check whether your category publishes a rate at all.** Today's release publishes 85 of 157 observed segments. The other 72 are collecting, which is not the same as empty. A segment with no published rate can still carry more observed answer runs than a published one. Do not read an absent number as zero demand.

**Three. Place yourself on the prominence ladder before choosing a lever.** If you are a category leader already appearing in most retrievals, more visibility spend buys little and differentiation buys more. If you are a specialist with roughly even odds of never appearing, the first job is minimum viable presence in one engine, not a broad program.

**Four. Fund the classes your category actually cites, in proportion.** In a vendor-led category, structured documentation and comparison pages are citation assets and deserve real budget. In an editorial-led category, outlet selection is the decision, and the long tail means one right placement beats five wrong ones. In every category, community and platform surfaces are few, concentrated, and not purchasable.

**Five. Re-read on a fixed cadence and hold denominators separate.** GEO measurement [varies across runs, prompts and time](https://arxiv.org/abs/2604.07585), so a single reading is not a baseline. The index refreshes daily and states its release id, window and run counts on every segment, which is what makes one reading comparable to the next.

## What this evidence does not establish

I would rather lose the argument than overstate it, so here are the limits in plain terms.

This is observed citation composition, not attribution. Nothing here shows that spending on a source class causes citations, and no honest reading of it produces a cost per citation. Three in five cited runs sit in an unclassified bucket, so every named share describes 39.5% of the map. The per-category figures use a top-100-domain denominator, not the full category. Eight of 25 categories have no published stratum yet. And citation is not the end of the mechanism: a [602-prompt analysis](https://arxiv.org/abs/2604.25707) separates being cited from being absorbed, where a source's evidence actually shapes the answer, and depth of absorption is not visible in a share table at all.

The engines publish their own retrieval behavior, and it is worth reading directly rather than through a vendor summary: [Perplexity's search guide](https://docs.perplexity.ai/guides/search-guide), [Google's AI features documentation](https://developers.google.com/search/docs/appearance/ai-features) and its [AI Mode announcement](https://blog.google/products/search/google-search-ai-mode-update/), and [Anthropic's web search tool](https://docs.anthropic.com/en/docs/agents-and-tools/tool-use/web-search-tool). The IETF also has an informational draft describing an [11-stage AI visibility lifecycle](https://datatracker.ietf.org/doc/draft-lynch-ai-visibility-lifecycle) from crawling through visible placement, which is a useful reminder that being crawlable is stage one of eleven.

The audience is real either way. [Forrester's read on generative AI and consumers](https://www.forrester.com/blogs/the-state-of-genai-and-consumers-for-2026) and [Pew's measurement of ChatGPT use among US adults](https://www.pewresearch.org/short-reads/2025/06/25/34-of-us-adults-have-used-chatgpt-about-double-the-share-in-2023/) both show the behavior is mainstream. The question is allocation, not whether to participate.

## Where this leaves the money

The uncomfortable implication of the table is that most AI visibility budgets are set by channel habit. Content teams propose content. PR teams propose PR. Tooling budgets grow because tools are easy to buy. None of those proposals is indexed to where citations land in that company's actual category, because until the index was public there was nothing to index them to.

The correction is cheap. Read your category's source mix, put yourself on the prominence ladder, fund the classes that carry citations in your market, and hold the denominators apart so next quarter's reading means something. That is a planning exercise, not a purchase.

The deeper point is about what kind of asset survives. Owned content is a necessary condition and a weak differentiator, because everyone can produce it and the engines discount it accordingly. What compounds is independent corroboration: other organizations, in the source classes your category cites, confirming the same claims. That is the discipline of [Machine Relations](https://machinerelations.ai), earning citation and recommendation through third-party credibility rather than through volume, and the index is simply the first public instrument that shows where the corroboration has to come from.

If you want your own category read against the competitive set rather than the portfolio average, a [visibility audit](https://app.authoritytech.io/visibility-audit) maps where your brand is cited and where the source ecosystem around it is thin.

## FAQ

### How should I prioritize an AI visibility budget?

Allocate against your category's observed source-class mix rather than the portfolio average. In release `mri_score_v2.0+2026-09-18+8fa38e54dd0a`, editorial publications carry 12.0% of role-attributed cited runs, vendor-owned sources 10.05% and wire distribution 0.20%, but the leading class flips by category: vendor-owned leads cybersecurity at 32.3% and enterprise software at 34.9%, while editorial leads healthcare services at 26.5%. Read your category first, then fund the classes that carry citations there.

### Is press-release distribution worth funding for AI citations?

Not on this evidence. Wire and press-release distribution is the smallest named source class in the index: nine domains, 226 cited runs, 0.20% of role-attributed cited runs across the 2026-05-10 to 2026-09-18 window. It remains useful for disclosure, syndication and creating a public record, and it scores 25.1 cited runs per domain because so few domains exist. Funding it as an AI citation strategy is not supported.

### Why does my category show no citation rate?

Because it is collecting, not empty. Today's release publishes rates for 85 of 157 observed segments, where a segment is a category paired with a buyer question shape. The remaining 72 are still accumulating observations, and a collecting segment can carry more observed answer runs than a published one. Treating an unavailable rate as a zero is the most expensive misreading available in this data.

### Does more content improve AI visibility?

Not reliably on its own. A study of [112 startups across 2,240 queries](https://arxiv.org/abs/2601.00912) found that optimizing site content for AI showed no correlation with discovery rates, while referring domains and community presence did correlate with Perplexity visibility. Owned content is necessary for an engine to understand what you do. Independent corroboration is what moves whether you get recommended.

### How often should an AI visibility budget be reviewed?

Quarterly for allocation, with monthly readings of your category's segment. [GEO measurement varies across runs, prompts and time](https://arxiv.org/abs/2604.07585), so a single observation is not a baseline. Compare like releases: the index states its release id, observation window and run counts on every segment, which is what makes two readings comparable.

## Links

- [Blog Index](https://authoritytech.io/blog.md)
- [Home](https://authoritytech.io/index.md)