Keep the AI visibility work that happens on your own site in-house, use a tool to watch the answers, and bring in an outside partner only for sources you do not control. In the Machine Relations Index release of October 20, 2026, answers to B2B buyer questions cited company-owned sites at least 71 times and third-party sources at least 114 times per 100 answer runs. Both lanes carry citations, and a different team controls each.
The question for a large company is rarely "agency or no agency". It is who can change which source. The company's team can change its site, crawl rules and measurement. Editors, analysts, review sites and communities it does not run decide the rest, and reaching them takes relationships that a corporate comms team may hold or an outside partner can bring. This page sizes each lane by question shape and category, then gives a decision table and a 90-day split. CMO budget and headcount sit on Christian Lehman's site.
It counts which sources six answer engines cite, so it can say how much of the cited pool a team can reach on its own site and how much sits elsewhere. The Machine Relations Index (MRI), published by Machine Relations Research, covers 28,201 answer runs from ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity between May 10 and October 20, 2026 (release mri_score_v2.1+2026-10-20+bf82d9d98978). A segment is one category paired with one buyer question shape. This page uses the 62 published segments in the 14 B2B categories for six shapes: best tools, how buyers choose, is it worth it, problem-first research, top lists and comparisons. That is 11,220 runs and 68,540 citations (a citation is one domain cited in one run), built on 434 distinct questions repeated across engines and dates.
GEO and AEO describe the on-site and answer-format work in the first lane; Machine Relations is the wider discipline that adds the third-party lanes, so the split below is by source class, not acronym. The Index assigns a class only on evidence, so most cited domains publish as "Other observed source". Every named class is a floor and Other is a ceiling (the correction record logs the gap), and the floors are not equally complete.
The team can change company-owned sites, at least 71.3 citations per 100 runs, and can only influence the other named classes, which total at least 114.4. The first class is the Index's "vendor-owned source": the sites companies run for their own products, every company's and not only yours, so one company controls one domain inside it. The "who can change it" column is this page's assignment, not an Index field.
| Source class (Index label) | Who can change it | Citations per 100 runs (floor) | Domains cited | Largest domain's share of its class |
|---|---|---|---|---|
| Company-owned sites (vendor-owned) | Web, product, docs and legal teams, for the company's own domain | 71.3 | 641 | microsoft.com 3.6% |
| Editorial publications | Whoever holds editor and journalist relationships | 28.1 | 360 | techradar.com 8.7% |
| Community and social platforms | Any employee, inside each platform's rules | 27.6 | 30 | reddit.com 29.0% |
| Market and company databases | Product marketing and customer teams, for profiles the company can edit | 21.1 | 380 | g2.com 12.2% |
| Academic and government sources | Research and engineering teams that publish | 19.8 | 248 | arxiv.org 31.6% |
| Analyst and consulting research | Analyst-relations team | 8.2 | 95 | gartner.com 17.4% |
| Search or media platforms | The channel owner, for the company's own channel | 8.0 | 20 | youtube.com 87.9% |
| Wire and press-release distribution | Anyone who pays for distribution | 1.6 | 17 | prnewswire.com 32.1% |
| Other observed source (unclassified) | Not known | 69.6% of all citations (ceiling) | not counted | not counted |
Concentration is the largest domain's share of its own class, never a top-N share. Company-owned sites and editorial publications have no incumbent: 641 and 360 domains share them, so each gain is one domain or one outlet at a time. Community and video platforms are head-heavy: a few domains hold most of each class, so no outlet list applies there.
Mostly at the edges: company-owned sites are the larger classified lane in all six shapes, from 51.4 per 100 runs for problem-first questions to 87.5 for best-tools questions. Read the first three value columns as floors and the last as a ceiling.
| Question shape | Segments / runs / distinct questions | Company-owned | Editorial | All other named third-party | Other (share of citations) |
|---|---|---|---|---|---|
| Best tools | 10 / 1,844 / 70 | 87.5 | 35.0 | 110.7 | 68.1% |
| How buyers choose | 10 / 1,855 / 70 | 62.1 | 24.0 | 77.1 | 72.0% |
| Is it worth it | 11 / 1,965 / 77 | 74.1 | 26.1 | 70.1 | 70.9% |
| Problem-first research | 11 / 1,951 / 77 | 51.4 | 27.3 | 72.1 | 68.9% |
| Top lists | 10 / 1,785 / 70 | 79.3 | 29.7 | 91.9 | 72.8% |
| Comparisons | 10 / 1,820 / 70 | 75.0 | 26.6 | 98.3 | 63.9% |
| All six shapes | 62 / 11,220 / 434 | 71.3 | 28.1 | 86.3 | 69.6% |
Values are citations per 100 answer runs. Problem-first is the shape a prompt such as "our Google traffic is flat, sign-ups are dropping and buyers say they found competitors through AI, where do we start" takes. It has the lowest company-owned floor and the highest community floor (31.4 per 100 runs, above editorial's 27.3), so more owners share that answer than a best-tools answer.
Company-owned sites are the larger lane in seven of the eight readable categories, from 1.9 times the editorial lane in HR & Talent to 4.0 times in Fintech, and the smaller lane in Deep Tech & Hardware. Readable means all six shapes published and Other under 80%; where Other passes 85% the named classes say almost nothing.
| Category | Published segments / runs | Company-owned | Editorial | All other named third-party | Other |
|---|---|---|---|---|---|
| Cybersecurity | 6 / 1,202 | 128.6 | 36.5 | 84.0 | 61.4% |
| AI Visibility & GEO | 6 / 1,405 | 104.6 | 31.1 | 92.3 | 70.3% |
| AI Security & Privacy | 6 / 1,430 | 98.5 | 30.3 | 126.1 | 60.1% |
| HR & Talent | 6 / 1,008 | 93.5 | 48.5 | 61.4 | 66.0% |
| Fintech | 6 / 1,097 | 70.4 | 17.5 | 57.4 | 74.0% |
| Enterprise Software | 6 / 1,070 | 66.6 | 31.8 | 80.5 | 69.0% |
| AI Infrastructure | 6 / 1,326 | 62.1 | 16.0 | 123.0 | 66.0% |
| Deep Tech & Hardware | 6 / 1,135 | 10.6 | 30.4 | 100.9 | 75.8% |
| Industrial | 6 / 714 | 0.7 | 16.0 | 38.0 | 90.1% |
| Logistics & Freight | 6 / 630 | 3.2 | 19.0 | 54.3 | 85.3% |
Industrial and Logistics & Freight are not readable. Martech & Advertising has two published buyer-shape segments (203 runs), and Professional Services, Sales, GTM & CRM and VC & Private Equity have none, so the Index offers no reading for those four. In the eight readable categories the editorial floor runs from 16.0 to 48.5 per 100 runs, so the third-party lane is never empty. HR & Talent has the highest editorial floor and the most concentrated: People Managing People holds 140 of its 489 editorial citations (28.6%), so there the outlet is the decision. Deep Tech & Hardware is led by academic and government sources (49.2 per 100 runs), which research staff publish into. In AI Visibility & GEO, where buyers ask this very question, company-owned sites are at least 104.6 per 100 runs, so engines' answers about hiring and tooling lean on sellers' own sites; read them for their evidence.
Pooled figures point the wrong way for a B2B company. Across all 24 categories the six shapes give editorial 52.4 and company-owned sites 43.4 per 100 runs, the same direction as the October 17 release, where editorial publications took 29.9% and company websites 25.6% of role-classified buyer-question citations (earned versus owned page). Split by buyer, the 14 B2B categories read 71.3 against 28.1 and the ten consumer and mixed categories 12.9 against 79.0. A University of Toronto preprint found AI search skewed to earned over brand-owned sources, but its regional and vertical experiment used 1,000 consumer ranking prompts and counted review and comparison sites as earned. It does not tell a B2B team where to start.
A tool should own repeated measurement of a fixed prompt panel, because one check, by software or by an agency reading answers by hand, is one draw. In the Index's question panel (980 questions, 22,620 runs), two random halves of the same question shared 46.1% of their cited volume, a reworded question 39.4% and a broader or narrower one 23.0% (Machine Relations Research). A preprint, not yet peer reviewed, reaches the same practical point: answers vary across runs, prompts and time, so visibility needs repeated measurement.
Tool coverage and data sources are the vendor's own. Semrush's documentation, read October 20, 2026, describes daily tracking of a custom prompt set across ChatGPT Search, Google AI Mode and Gemini and lists the domains cited for each tracked prompt, while its overview reports draw on a separate prompt database of responses captured from real requests. Check any tool's engines and data source against your buyers' and ask how it reports failed runs: the Index's manifest records Gemini returning a citation in 2,153 of 6,384 returned runs. Nine vendors' disclosures are audited in AI Visibility Tools vs. an Agency.
The in-house team owns whatever only the site owner can change: access, the facts on company pages, first-party measurement accounts and custody of the raw answer data. Four cases where in-house wins outright:
Keep prompts, dates, engines, answers and source URLs in company storage whatever the vendor. The handover checklist lists what an export must carry; Jaxon Parrott's build-versus-buy note states the founder boundary.
An outside partner earns its place in the third-party lane where the company holds no relationship, and it should be judged by the outlets on your category's leaderboard, never by a citation promise. Disclosure: AuthorityTech is an earned-media agency that sells this lane, so it has an interest in how this section reads; every figure here can be re-derived from the public artifact. Editorial publications are 28.1 per 100 runs across 360 outlets, the largest at 8.7%, so which outlet is the work. Press-release distribution is 1.6 per 100 runs. The Index records what engines cited, not what caused it, so no placement or page can be promised a citation.
AuthorityTech publishes no rate card. Scope and fees are set on a strategy call, there are no retainers, and placements are paid after publication with funds held in escrow until a placement is live (authoritytech.io/#pricing). For which kind of firm to hire first, read GEO Agency vs SEO Agency vs PR Firm, and use the RFP questions before signing anyone.
Assign each task to whoever can change its source, and ask every vendor for the evidence that task needs.
| Task | Owner | What the evidence says | What to ask a vendor |
|---|---|---|---|
| Crawl and index access | In-house web team | Google: indexed and snippet-eligible, nothing more; each engine documents its own bots | Who has write access to robots.txt and CDN rules; keep it with your team |
| Prompt panel and repeated measurement | Tool, with an in-house panel | A rerun shares 46.1% of cited volume; the Index uses 7 questions per segment | Which engines, how many runs per prompt, how failed runs are reported, which raw answers export |
| Company-owned pages: answers, docs, comparisons | In-house product and web teams; an agency may draft | At least 71.3 per 100 buyer runs, 641 domains, largest 3.6% | Who writes, who approves facts, who publishes |
| Editorial placements | Comms team or outside partner | At least 28.1 per 100 runs, 360 outlets, largest 8.7% | Which outlets on my category leaderboard, with live URLs; how placement, mention, host citation and exact-URL citation are reported apart |
| Review sites, analyst and research profiles | Product marketing and analyst relations | Databases at least 21.1, academic 19.8, analyst 8.2 per 100 runs | Who owns each profile and how corrections are made |
| Community and video platforms | Company staff inside platform rules | Reddit 29.0% and YouTube 87.9% of their classes | Which accounts and platform rules apply; decline any promise of placement |
| Referrals and pipeline | In-house analytics | The Index counts citations, not visits or revenue | How the vendor's report joins your CRM, if at all |
Spend weeks 1 to 2 on measurement and access in-house, weeks 3 to 6 on company-owned answers, and weeks 7 to 12 on the third-party lane, then re-read the same panel.
| Weeks | In-house | Tool | Outside partner |
|---|---|---|---|
| 1 to 2 | Write a panel of seven questions per shape. Audit crawl rules and indexing. Verify Search Console and Bing AI Performance. | Baseline the panel on your buyers' engines, repeated on different days. | None yet. Read your category's leaderboard on the Index. |
| 3 to 6 | Fix access defects. Publish or revise company-owned pages that answer each shape, facts approved by their owners. | Weekly read of cited classes and domains per question; keep raw exports. | Compare the leaderboard's outlets with the comms team's relationships; scope a partner for the outlets nobody on staff reaches. |
| 7 to 12 | Re-run the panel and compare class mix with weeks 1 and 2. Decide whether a company-owned gap remains. | Export all answers into company storage. | Placements go live per scope, reported as placement, mention, host citation and exact-URL citation. |
The Index shows no cost per citation for any class, and a page, a placement or an access change may precede a citation without causing it. A class share describes where engines cited in this window and is not an ROI figure.
Split it by who can change the source. Keep the company-owned site, access rules, facts and data in-house; those were at least 71.3 citations per 100 B2B buyer-question runs in the October 20, 2026 Index release. Hire outside for third-party sources where no one on staff holds the relationship: editorial publications were at least 28.1 per 100 runs.
Buy the tool for repeated measurement and keep the exports. A rerun of the same question shared only 46.1% of cited volume in the Index's panel, so a hand check is one draw. Pay an agency for work a tool cannot do, which is changing what engines cite.
SEO covers access: Google says AI Overviews and AI Mode need an indexed, snippet-eligible page and nothing more technical. It does not cover the third-party lane: Moz found only 12% of AI Mode citations matched organic URLs for the same query across nearly 40,000 queries. Staffing is on Christian Lehman's site.
Start with a fixed panel across the six question shapes, an access audit and a read of your category's class mix. For problem-first questions the Index shows company-owned sites at least 51.4, community platforms 31.4 and editorial publications 27.3 per 100 runs, so no single owner covers the answer.
AuthorityTech publishes no rate card and sets scope and fees on a strategy call. It has no retainers; placements are paid after publication, with funds held in escrow until a placement is live.
It is the benchmark of which sources AI answer engines cite, published by Machine Relations Research at machinerelations.ai and updated daily. Machine Relations is the discipline Jaxon Parrott coined in 2024; AuthorityTech, which he founded in 2018, is the first Machine Relations agency. The Index is Machine Relations Research's instrument, not an AuthorityTech tool.
Method note. Figures are the public Index artifact for release mri_score_v2.1+2026-10-20+bf82d9d98978 (window May 10 to October 20, 2026, 157 days observed; 55,217,722 bytes, sha256 beginning bf82d9d98978, matching the release manifest). Each class figure sums the cited-run counts in every domain's per-segment rows, grouped by the domain's source role, over the 62 published segments in the 14 B2B categories for six shapes: 11,220 runs, 68,540 citations (one domain cited in one run) and 434 distinct questions. News-driven segments and the 22 B2B segments below the evidence floor (782 runs) are left out; all 89 B2B segments total 14,928 runs and 95,630 citations. The 14 B2B categories are AI Infrastructure, AI Security & Privacy, AI Visibility & GEO, Cybersecurity, Enterprise Software, Fintech, HR & Talent, Industrial, Logistics & Freight, Martech & Advertising, Professional Services, Sales, GTM & CRM, VC & Private Equity and Deep Tech & Hardware; the consumer comparison uses the other ten (54 segments, 10,260 runs). "Per 100 runs" is class citations divided by runs, times 100. It can pass 100 because one run can cite several domains, and it is not the share of runs citing the class, which the public artifact cannot give because it holds no run-level rows. Named classes are floors and Other observed source is a ceiling; "Other" in a table is its share of that row's citations. Concentration is the largest domain's citations divided by its class's citations over the same segments. Re-query each segment on its Index page, for example the Cybersecurity how-buyers-choose page, which lists microsoft.com at 42 of 218 runs, or in the public JSON artifact.