The Visibility Method

The method answers one question: can this page be fetched and quoted, and does a live answer name the campaign, place it, cite it, and describe it in a way that matches the page? A fetchable page that is described wrongly is not a strong result. A flattering answer that the page cannot support is not a strong result either. Both kinds of evidence stay in the same total, and each measure stays on its own line.

1. Read. Fetch the submitted URL, robots.txt, llms.txt, and sitemap.xml. Score ten page measures. Their maxima are 7, 8, 4, 6, 8, 12, 4, 2, 1, and 3. They sum to 55.
2. Ask. Send four prompts to one engine, Grok, with web search. The prompts use the campaign name, type, city, and state. Score six answer measures: mention rate 9, answer position 6, share of voice 6, citation sources 12, description accuracy 8, and sentiment 4. They sum to 45. A high mention rate does not cancel an inaccurate description or a negative tone.
3. Add. When the ask step scores at least one measure, add the two integers. The result is out of 100. When the ask step is not scored, publish the page measures only, out of 55. Do not fill the missing 45 points with zero.

The four platform numbers are a reweight of the read step. They are not a fourth step, and they do not query those products. The rules for each measure are below. This page is the spec for the free check. It is not the managed Election GEO assessment. A higher total means more of this evidence was present. It is not a forecast of traffic, donations, or votes.

The read step requests four public addresses:

• The submitted URL, followed through up to four redirects. The score uses the final URL.
• /robots.txt on the final origin.
• /llms.txt on the final origin.
• /sitemap.xml on the final origin. A missing robots, llms, or sitemap file does not fail the scan.

The request uses the user agent OdysseyCampaignsVisibilityScan/1.0 and a link back to this site’s check page. Each page request waits up to 12 seconds and accepts a body of up to 1.5 MB.

The ask step is the same run, not a second product. When an engine key is configured, it sends the four prompts to the xAI Responses API with web search, two at a time. Each prompt waits up to 80 seconds. If the key is missing, or every prompt fails, the method still returns the read step and says the answer measures were not scored. A prompt that returns no text is named in the result and is left out of the ratios. Its points are not given to the signals that did return.

Only http and https URLs are accepted. The host must resolve to a public address. If the local resolver fails, or returns a private address, the check asks Cloudflare DNS for a public A record and connects to that address with the original host name. Local hosts, private addresses, and URLs with a username or password are refused. A page over 1.5 MB, more than four redirects, a non-HTML document that returns HTTP 200, or a timeout ends the scan with an error instead of a low score.

The scan route allows eight checks per IP address in a rolling minute. The limit is kept in memory on that server instance.

Each measure produces a ratio from 0 to 1. The ratio is multiplied by that measure’s maximum. The ten page measures are turned into one integer inside 55 points. The six answer measures are turned into one integer inside 45 points. The method adds those two integers. It does not throw all sixteen lines into one rounding pool. The page maxima are 7, 8, 4, 6, 8, 12, 4, 2, 1, and 3. The answer maxima are 9, 6, 6, 12, 8, and 4.

Inside each step, the displayed points are integers and they sum to that step’s integer:

1. Add the raw products in that step.
2. Round that sum to the nearest integer and keep it inside the step’s cap, 55 or 45.
3. Start each measure at the whole-number part of its own product.
4. Give the remaining points, one at a time, to the measures with the largest fractional parts.

A measure can show fewer points than a simple rounding of its own product when another measure in the same step had a larger fraction. If an answer measure cannot be scored because the answers do not support it, it is left out and the remaining answer weights are scaled to 45. A measured zero stays. If a prompt returns no text, the measures that were scored keep their own maxima and those missing points are not reassigned. If no answer measure can be scored, the method total is the page integer, out of 55.

Each platform column runs the same rounding method on the read step, with its own weights. The weights in a column sum to 55. They are a published emphasis for that column, not a measurement from that product, and they are not the six answer measures.

Crawl access is the one exception. A platform score uses that platform’s own allow-share, not the averaged crawl ratio used in the total. If robots.txt is missing, each platform allow-share stays at 1, so the platform crawl line is full credit, while the total crawl line uses 0.75.

The large number is the Visibility Method total. It is out of 100 when the ask step is scored: the page integer plus the answer integer. It is out of 55 when the ask step is not scored, so a missing sample is not stored as a zero. Under it, each of the sixteen measures shows its own points. The six answer measures stay separate so a high mention rate can sit next to a wrong description or a negative tone.

The four platform numbers are the read step, reweighted, out of 55. They can sit away from each other because their weights differ and because a blocked crawler changes only that platform’s crawl line. They are not four answer audits.

Use a low line as a fix list for the page that was scanned: publish the missing file, add the missing heading and answer, add JSON-LD, or put the name, street, and phone on the page. The finding sentence is the check’s own description of that URL. It is not a recommendation generated by a model.

On October 14, 2026 the same ten homepages were scored with this check and with the open-source GeoScore audit at geoscoreapp.pages.dev. GeoScore splits a technical SEO score from a GEO score. This check’s old total tracked the SEO number. The GEO number stayed much lower on the campaign and product sites.

None of the ten robots.txt files blocked an AI crawler, so crawler access did not explain the gap. llms.txt was already present on several sites whose GEO score was still under 30, including sites that scored below a site with no llms.txt. Stripe, the one site with a Wikipedia article, was the high GEO score. The repeated GeoScore findings on the other nine were no Wikipedia page, no Wikidata entity, and no social profile on the homepage.

That comparison is why this is one method. The old total tracked GeoScore’s SEO column and left the GEO gap unpublished. In that GeoScore audit the overall figure mixed its SEO score at about 55 percent with its GEO score at about 45 percent. The method keeps both: 55 points on the read step and 45 on the ask step, added into one total. Inside the read step, discovery files are 1 point and structured data is 12, because llms.txt did not separate the GEO column and an entity link did. Citation sources take 12 of the 45 ask-step points for the same reason. Sentiment is 4 points of its own, not a footnote on accuracy.

These six measures are the ask step. They belong to the same total as the ten page measures. The read step says whether a page can be fetched and quoted. The ask step says whether an answer names the campaign. Each measure is still shown on its own. A high mention rate with an inaccurate description, or with a negative tone, is two lines in one result, and both can lower the total.

The sample is four prompts on one engine, Grok, with web search, when an engine key is configured for the server. The recorded context is the prompt, the raw answer, the engine and surface, the date and time, country United States, account state “api”, the model name the API returns, and the visible citation URLs. This is not a logged-out consumer session on ChatGPT, Perplexity, Gemini, or Google, and it is not a market-wide rate. If the key is missing or the sample returns no text, each answer measure stays blank and the total is the page measures only.

Three prompts leave the campaign name out. They cover finding options and comparing options, for a voter or a buyer, in the city and state from the form. The fourth prompt names the campaign and asks for background, office, and contact information. That fourth prompt is for description accuracy only. It is not eligible for mention rate, answer position, or share of voice. A discovery prompt that already contains the brand is also ineligible, so the brand name cannot inflate mention rate.

• Mention rate is the eligible answers that name the brand, divided by eligible answers. A direct name, an abbreviation, or a product name counts. A fuzzy reference, such as the domain with no name, is counted beside the rate and does not increase it.
• Answer position is the first explicit mention of the brand. The buckets are first recommendation, inside the first three named options, a longer list, or a caveat rather than a recommendation. An ordered list uses the list order. With no list, the check uses the order the names appear in the text.
• Competitor share of voice is this brand’s mentions divided by mentions of every name tracked in that same buyer, job, and stage. Names supplied as competitors count, and so do other names an answer lists. The check does not add those groups into one share.
• Citation sources record the URL the answer shows and its root domain. Each domain is classed as your website, a review or directory, an editorial publication, a community or forum, a competitor website, or another third party. A domain is counted once in an answer, even when several visible URLs share it. Own-domain coverage is eligible answers that cite your domain, divided by eligible answers. An answer with no visible URL is “no visible citation.” It is not evidence that the model used no source. URLs a tool opened but did not show are stored apart from citations.
• Description accuracy is graded only where the answer names the brand, against the phone on the page and the city, state, and office from the check. The grades are accurate, partly accurate, inaccurate, and unverifiable. Accurate scores 1, partly accurate scores 0.5, and inaccurate scores 0. Unverifiable answers are left out of the points.
• Sentiment is the tone of those named answers: recommended, neutral, cautionary, or negative. Recommended scores 1, neutral 0.6, cautionary 0.3, and negative 0. The supporting sentence is kept with the raw answer. Sentiment does not change the accuracy grade.

Answer position turns the buckets into points: first recommendation 1, inside the first three 0.7, a longer list 0.35, and a caveat 0.2. Share of voice uses the mean of the prompt-group shares. The groups stay listed, and they are not added into one raw count. Own-domain citation coverage is the citation ratio.

A line under a signal still names a next step. Mention rate under 50 points to category relevance and third-party presence. Fewer than half of the named answers in the first recommendation or the first three points to differentiation and comparison evidence. Share of voice under 30 inside a prompt group points to the sources that support the names leading that group. Own-domain coverage under 25 points to source accessibility and external references. An inaccurate or partly accurate description points to the official facts on the page and to stale third-party profiles. A sample that is mostly cautionary or negative points to the pages that frame the brand that way.

The scan request sends the URL, the campaign or committee name, the type, the city, and the state into one method run. The read step uses the public page. The ask step uses the name and place for the prompts and for the sentiment and accuracy grades. If the lead record fails, the score still stays on the page.

When the lead record succeeds, the page says the work email was recorded and that a person can follow up. The score is not emailed. The server writes a log line with the campaign name, email, URL, category, city, state, optional contact name, and the time. No phone number is collected.

Run the free AI visibility check on the campaign site you want scored. For the managed practice, see Election GEO.