Machine Relations

Is an AI Visibility Tracker Worth Paying For? Only If It Shows Response Rate, Runs and Model Version

Returned answers carried a citation 33.7% to 99.98% of the time across six AI engines. What to check before you pay for an AI visibility tracker.

Jaxon ParrottOct 20, 2026

Pay for an AI visibility tracker only if it shows, per engine, how many answers carried a citation, how many runs sit behind each prompt and which model version answered. In the Machine Relations Index release of October 20, 2026, 33.7% to 99.98% of returned answers carried a citation, depending on the engine, and a rerun of one question shared 46.1% of its cited-source volume.

A tracker that hides those numbers gives you a score that can move when an engine goes quiet, a model is swapped or a prompt is reworded, with no change in how buyers see your brand. One that shows them is a measuring instrument, worth a subscription once you have a stable prompt list and an owner for the gaps it exposes. The Machine Relations Index (MRI), published by Machine Relations Research, samples six engines and prints its own collection health per engine, so the numbers behind this decision can be read in full and re-queried.

Key Takeaways

Engines respond at very different rates. Returned answers carried a citation 99.98% of the time on Perplexity, 88.5% on ChatGPT, 81.1% on Claude, 77.8% on Google AI Mode, 66.0% on Google AI Overviews and 33.7% on Gemini.
Collection alone moves the score. A brand cited in half of the answers that cite anything scores 50.0% on Perplexity and 16.9% on Gemini over returned runs. That is arithmetic, not a measured brand.
The model behind an engine label changes. Gemini's runs carry six model ids in one window and ChatGPT's three; 12.7% of runs carry none.
One run is a draw. A rerun shares 46.1% of cited-source volume, a reworded prompt 39.4%, a broader or narrower one 23.0%.
Test for live collection on the per-run record, not the dashboard: timestamp, model id, search flag, and source lists that differ run to run.
Pay when volume or history outgrows a hand-run panel, not while the prompt list is unsettled or no one owns the gaps.

Can a tracker's number mean the same thing on every engine?

No: the Index's six engines returned a citation on 33.7% to 99.98% of their answers, so a blended figure mixes engines observed on different days, at different response rates, through different model versions. Release mri_score_v2.1+2026-10-20+bf82d9d98978 covers 28,201 answer runs from May 10 to October 20, 2026; a run is one question asked of one engine on one day. The release manifest lists each engine's days answered, returned runs and returned runs that produced a citation. The Index's own denominator also removes 948 runs on 81 engine-days when an engine ran at least five times and returned no eligible citation anywhere in the corpus.

Engine	Days observed (of 157)	Returned runs	Runs with a citation	Share, returned-run basis	Share, Index denominator	Model ids that answered
Perplexity	157	6,510	6,509	99.98%	99.98% (of 6,510)	1 (243 runs with none)
ChatGPT	153	6,399	5,661	88.5%	88.5% (of 6,399)	3 (216 runs with none)
Claude	156	6,503	5,274	81.1%	81.1% (of 6,503)	1 (216 runs with none)
Google AI Mode	115	1,966	1,530	77.8%	86.2% (of 1,775)	None exposed
Google AI Overviews	98	1,387	915	66.0%	81.3% (of 1,125)	None exposed
Gemini	135	6,384	2,153	33.7%	36.6% (of 5,889)	6

The response rate belongs to the collection, not to any brand: across all six engines, 22,042 of 28,201 Index runs (78.2%) carried a citation. Gemini supplied 20.9% of the runs but 9.8% of those with a citation, and the release's own coverage statement discloses it, with both Google surfaces, as observed on under 90% of days. The denominator rule matters too: Google AI Overviews reads 66.0% or 81.3% depending on whether 262 excluded runs stay in, and the Index lists every excluded engine-day so a reader can choose. Google's documentation says AI Overviews are shown only when its systems judge them additive to classic Search and "often don't trigger".

The arithmetic follows. A brand cited in half of the answers that cite anything scores 50.0% on Perplexity, 44.2% on ChatGPT, 40.6% on Claude, 38.9% on Google AI Mode, 33.0% on Google AI Overviews and 16.9% on Gemini when counted over returned runs, and a blended number inherits that spread. The count of engines a tracker lists is not the feature; the response rate and days observed beside each engine are. Per-engine observation coverage has the day-by-day roster.

The model behind an engine label also moves. Each run carries the id the provider returned: Gemini's 5,889 Index runs list six (gemini-3.8-flash answered 3,333; gemini-2.0-flash, 2.5, 3.5, 3.6 and 3.7 flash the other 2,556) and ChatGPT's 6,399 list three (gpt-4.1-mini 3,830, gpt-4o-mini-search-preview 1,969, gpt-4o-mini 384). The two Google surfaces expose none, so 3,575 runs (12.7%) carry no id. A trend line that crosses a model swap can show a step the brand did not cause.

How many runs does a tracked prompt need before you act on it?

Treat a prompt's source list as a draw until a batch of runs stands behind it: in the Index a rerun shares 46.1% of cited volume, a reworded prompt 39.4% and a broader or narrower one 23.0%. The rewording study splits each question's runs into random halves (980 questions, 22,620 runs, July 6 to October 20, 2026) and measures how much cited volume two halves share.

Compared with the question's own half	Shared cited volume	What it means for a tracker
Other half of the same question (a rerun)	46.1%	Same prompt, same engines: over half of the cited-source volume still differs
Half of a reworded question	39.4%	One wording is one draw of the wording as well as of the runs; it keeps 77 to 88% of the rerun overlap
Half of a broader or narrower question	23.0%	Another question, not a rerun; it keeps 44 to 49% of the rerun overlap
Two single runs of one question	14.3%	31.6% on the same engine across days, 9.0% across engines

Repetition separates a source that is there from one that appeared once. Among 658 questions with at least 8 runs per half, a domain cited in one run of the first half was cited again in the other half 37.8% of the time (19,828 question-domain pairs); with two citing runs 64.7%, three to four 83.2%, five or more 95.6%. The sampling error on a rate follows the binomial distribution (NIST): at a 50% rate the 95% half-width, 1.96 times the square root of p(1-p)/n, is 31.0 points at 10 runs, 17.9 at 30, 9.8 at 100 and 5.0 at 384. That is a floor, since runs of one prompt are not independent draws and wording adds variance.

Scale is the limit. The Index runs 1,227 monitored prompts, 23.0 runs each across all engines, and publishes rates only for category-and-question-shape segments, the smallest holding 100 runs. A tracker printing a prompt-level rate from one daily run on one engine has about 30 runs a month behind it. The unrefereed April 2026 arXiv preprint Don't Measure Once reaches the same conclusion: answers vary across runs, prompts and time, so visibility is a distribution that needs repeated measurement.

How do you tell a live tracker from a prompt-replay dashboard?

Ask for the per-run record (timestamp, engine, resolved model id and whether the engine searched), then check that the source lists differ from run to run. "Live or cached" hides three designs, and each answers a different question. The wording below is from each vendor's own page, read on October 20, 2026.

Design	Question it answers	What a vendor page states
Scheduled runs of your prompts on the engines	What would the engines say to the prompts I chose, on the days measured?	Profound: prompts are "the queries Profound sends to answer engines daily". OtterlyAI: tracked prompts run daily on every engine it checks. Evertune: every prompt sampled 100 times per model.
A database of captured responses to real prompts	What did engines answer to common prompts, and when was each captured?	Semrush: "Prompt responses are captured from real requests and not via any APIs of LLMs", updated daily on a rolling basis.
First-party reports from the engine's operator	How often were my own pages cited in real answers?	Bing Webmaster Tools AI Performance reports citations across Copilot, AI summaries in Bing and select partners, and calls its grounding-query data "a sample".

A stored answer is not a defect; a stored answer sold as a fresh run is. Three checks separate them in a demo. First, take a prompt you choose and compare the vendor's run for today with your own run in the engine's interface: expect partial overlap, not identity and not zero, since two single same-engine runs shared 31.6% of cited volume in the Index. Second, ask for seven days of per-run records for one prompt: ten identical source lists in a row is a stored answer or an average, and you should ask which. Third, ask whether the record shows that the engine searched: OpenAI's documentation says the model can choose to search the web or not, and Anthropic's web search tool reports a web_search_requests count in the response usage block. Ahrefs separates pages "found in" an answer from pages cited in it.

What should a tracker show before you pay?

Five things, each backed by what the Index measured: runs per prompt, a response rate beside every engine, a per-run record, a named counting unit and the prompt set itself. The checklist pairs each with Index evidence and with what vendor pages state. It ranks no vendor and says nothing about a capability a page does not mention. AI Visibility Tools vs. an Agency holds the vendor-by-vendor disclosure audit.

Ask for	Why: what the Index measured	What vendor pages state (October 20, 2026)
1. Runs per prompt per engine, and cadence	A rerun shares 46.1% of cited volume; a source seen in one run reappears 37.8% of the time, in five or more runs 95.6%	Evertune: 100 runs per prompt per model, with a claimed margin of error of about one point overall. OtterlyAI: daily. Scrunch: rate charts show sample size, for example "33% (3 of 9 responses)".
2. Engines, with response rate and days observed beside each	33.7% to 99.98% of returned answers carried a citation; engines answered on 98 to 157 of 157 days	OtterlyAI lists ChatGPT, Google AI Overviews and AI Mode, Perplexity, Gemini, Microsoft Copilot and Claude as checked daily.
3. A per-run record: timestamp, resolved model id, search flag	Six Gemini ids and three ChatGPT ids in one window; same-engine runs on different days share 31.6%	Ahrefs: "found in" versus cited. Semrush: captured responses, not LLM API calls.
4. The counting unit and denominator behind each headline number	The Index counts a domain once per run; one engine reads 66.0% or 81.3% under two denominators	Peec: visibility is responses mentioning the brand over total responses. Profound: responses including the brand over responses including at least one brand. Scrunch: citation URLs, citing responses and all responses are three metrics.
5. The prompt set, its wordings and its version	A reworded prompt keeps 77 to 88% of the rerun overlap, a broader or narrower one 44 to 49%; the Index publishes a SHA-256 fingerprint of each of 147 question sets	Profound: prompts are grouped into topics and tags. Semrush: prompts come from clickstream data, grouped into topics, duplicates removed and phrasing simplified.

Item 4 is where two dashboards part ways on the same brand: Peec and Profound both name a visibility score, and the quoted definitions divide by different totals. How citation rate is calculated and the nine-platform methodology audit work through the denominators, and the methodology transparency benchmark covers 16 platforms. A vendor that supplies items 1 to 5 in writing has given you a contract you can audit. One that supplies a score without them has given you a number you cannot test.

When do you not need to pay for a tracker?

You do not need one while the prompt list is unsettled, while your question is about your own pages, while your buyers sit on one or two engines, or while no one owns the gaps a tracker would expose. You do need one when the same prompts must run on several engines, repeatedly, with history and exports, at a volume a person cannot log.

Your situation	Do this	Evidence
No stable prompt list yet	Run a hand panel first: 20 prompts, 3 engines, 5 runs each is 300 answers per cycle	300 answers give about 5.7 points of sampling error at a 50% rate; the Index's smallest published segment holds 100 runs
You want to know whether your own pages are cited	Read the first-party reports before paying	Search Console is a free Google service, and Google counts AI Overviews and AI Mode in its Performance report under the Web search type. Bing's AI Performance covers Copilot citations. Microsoft Clarity's citation dashboard reports page citations and share of authority, and Microsoft calls Clarity free.
Your buyers use one or two engines	Pay for those engines, not for an engine count	The response rate differs by engine (33.7% to 99.98%), so an engine count says little about the reading
No one owns the gaps	Do not pay yet	A tracker records what engines cite; the Index records outcomes and cannot show why
You need 100 or more prompts, several engines, daily history and exports	Pay, and require items 1 to 5	100 prompts on 4 engines once a day is 400 answers daily, beyond a hand log

The first-party reports measure your own site, not a competitor set or the prompts you choose, so they answer a narrower question than a panel does. The Index category page for AI Visibility & GEO shows which sources engines cite for tracker-choice questions at no cost. Disclosure: AuthorityTech is an earned-media agency. It sells placements in publications engines cite, not a tracker, so it has an interest in how this page reads. It publishes no rate card and sells no self-serve plan; 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 the placement is live.

Whichever tracker you choose, it reports what the engines cite; it does not change it. AuthorityTech is the first Machine Relations agency, and its work is earning the third-party coverage those engines draw on. Start your visibility audit →

What the Index cannot tell you

It does not test any tracker. The response rates belong to the Index's own prompts, provider access and dates; a vendor using different access can differ, which is why each vendor should publish its own.
Its panel is not your prompts. It runs research questions in 24 categories; wording effects on your customers' prompts are not measured.
Its unit is the cited domain. It does not record brand mentions in answer text, so none of these figures is a share of voice.
It shows outcomes, not causes. Why Gemini returned fewer citations is not in the release, and 12.7% of runs carry no model id.

Frequently asked questions

Do you need to pay for an AI visibility tracker at all?

Not at first. Run a hand panel of about 20 prompts on the engines your buyers use and read the free first-party reports from Google Search Console, Bing Webmaster Tools and Microsoft Clarity. Pay once the same prompts must run on several engines, repeatedly, with history and exports, at a volume you cannot log.

How do you tell whether an AI visibility tool queries live models or replays a cached snapshot?

Ask for seven days of per-run records for one prompt: timestamp, engine, resolved model id, whether the engine searched, and the cited sources. Live runs differ from each other; in the Machine Relations Index, single same-engine runs shared 31.6% of cited volume. Identical source lists across days indicate a stored answer or an average.

Why do two trackers report different numbers for the same brand?

They differ in prompts, runs, engines, model versions and denominators. Peec divides mentions by total responses; Profound divides by responses that include at least one brand. A reworded prompt alone shares 39.4% of cited-source volume with the original in the Index.

Is the Machine Relations Index an AI visibility tracker?

No. It is the benchmark of which sources AI answer engines cite, published by Machine Relations Research and updated daily. It samples the market's questions and does not monitor any brand's prompts. Machine Relations is the discipline Jaxon Parrott coined in 2024; AuthorityTech, which he founded in 2018, is the first Machine Relations agency.

Method note. Index figures are from the public artifact and release manifest for release mri_score_v2.1+2026-10-20+bf82d9d98978 (window May 10 to October 20, 2026, 157 observed days; artifact 55,217,722 bytes, sha256 beginning bf82d9d98978, matched to the manifest), which counts 28,201 runs, 1,227 monitored prompts and 205,362 citation events from six engines. Days, returned runs ("runs_total") and runs with a citation ("runs_observed") are the manifest's engine_roster; its coverage statement describes the latter as returned runs that produced a citation. The 948 excluded runs (Gemini 495, Google AI Mode 191, Google AI Overviews 262, on 81 engine-days) are the artifact's denominator_exclusion block, and returned runs less those equal the per-engine run counts summed across its question_context strata. The response rates, the 50% illustration, the run shares and the binomial half-widths are AuthorityTech's own calculations from those fields, not published Index figures. Model ids are summed from the strata's engine_model_runs field; 24,626 of 28,201 runs carry one. Rerun, reword, single-run and reappearance figures are read from the published rewording study, which uses a different panel and window. The artifact carries no prompt or answer text and this page uses none. Vendor and engine pages were read on October 20, 2026 and quoted as written; a claim about a vendor is a claim about its page, and a page that does not mention a capability is not evidence the product lacks it. Re-query the release in the public JSON artifact, the release manifest and the Index page.