Ready to Train the Bots

Schema markup is machine-readable code embedded in your website that tells AI and search engines exactly what your content represents. It uses a vocabulary called Schema.org, an open standard maintained by Google, Microsoft, Yahoo, and Yandex. The markup is added as a JSON-LD script block in your page's HTML head. A basic GovernmentOrganization tag for a county election office looks like this:

{
  "@context": "https://schema.org",
  "@type": "GovernmentOrganization",
  "name": "Larimer County Elections",
  "url": "https://www.larimer.gov/elections",
  "address": {
    "@type": "PostalAddress",
    "streetAddress": "200 W Oak St",
    "addressLocality": "Fort Collins",
    "addressRegion": "CO",
    "postalCode": "80521"
  },
  "telephone": "+1-970-498-7820",
  "areaServed": "Larimer County, Colorado"
}

The most useful schema types for election organizations

• GovernmentOrganization identifies your office as an official government entity with name, address, and jurisdiction.
• Event tags election dates, filing deadlines, public hearings, and candidate forums with structured date, time, and location data.
• FAQPage marks up your voter FAQ so AI can extract and cite specific answers directly.
• Person correctly identifies election officials, their roles, and contact information.
• Organization is commonly used by NGOs and commercial vendors to establish entity identity.

Schema markup helps AI systems interpret and attribute your content correctly. Microsoft has confirmed it helps its models. Broader effects on AI citations are not yet proven, but the markup is low-cost and aids readability and accessibility. When a voter asks ChatGPT or Perplexity when voter registration closes in their county, structured data helps ensure the AI cites your official page rather than a news article, advocacy site, or outdated archived page.

Web accessibility and AI-readability are approached from different angles, but content that is accessible to screen readers is content AI systems can also process accurately, because both depend on the same underlying structure.

Practically, this means adding descriptive alt text to all images, including any text visible within the image, and using a consistent heading hierarchy where H1 is the page title, H2 marks major sections, and H3 marks subsections, never skipping levels. PDFs should be text-based rather than scanned images; use Optical Character Recognition (OCR) where needed. Tag headings and lists properly. Tables need proper header elements so AI understands what each column represents. Link text should be descriptive: "2025 General Election Calendar" tells both a screen reader user and an AI crawler what the link leads to, while "click here" does neither.

Election officials interested in learning more about web accessibility may find value in the Section508.gov Content Library for additional guidance, resources, and information.

AI systems are pattern-matching engines. Inconsistency in how you refer to offices, procedures, dates, and officials creates confusion in what the AI learns, which produces inconsistency in what it repeats.

The fix is straightforward. Use the official, legal name of your jurisdiction consistently across every page. Write dates in full: "November 4, 2025" rather than "11/4" or "Nov. 4th." Spell out deadlines explicitly, as in "Registration closes at 5:00 p.m. Central Time on October 7, 2025." Expand jargon acronyms on first use, for example "Voter Identification Number (VIN)" before shortening to VIN. And once you pick a term for a concept, use it everywhere: do not alternate between "polling place," "polling location," and "voting site" as though they are interchangeable.

AI systems favor authoritative sources. For election offices, authority signals include operating on an official .gov domain where applicable, maintaining consistent Name, Address, and Phone (NAP) data across your website and partner directories, and earning inbound links from trusted organizations like state election offices, the National Conference of State Legislatures (NCSL), the National Association of Secretaries of State (NASS), the Cybersecurity and Infrastructure Security Agency (CISA), and the Electronic Registration Information Center (ERIC). Maintaining an active presence in authoritative directories like USA.gov and vote.gov helps, as does publishing regular content updates with visible timestamps. AI systems weight freshness.

For NGOs and commercial vendors, authority signals matter equally. If your organization publishes election information or supports officials, make sure your "About" page clearly describes your mission, relevant funding sources if applicable, and relationship to official election administration. AI systems increasingly evaluate source credibility, and an organizational identity that is ambiguous or hard to verify can result in your content being deprioritized or mischaracterized.

AI models learn from the web, and the web is messy. Election misinformation frequently enters AI training pipelines through sources that look credible but are not: old government pages with outdated deadlines, archived news articles citing superseded rules, unofficial clones of official sites, and social media posts scraped into training datasets.

Common sources of bad election data in AI systems

• Stale official pages with old election calendars, former officials' contact pages, and superseded procedures that remain indexed and crawlable.
• Unofficial clones like advocacy sites, watchdog sites, and local news archives that reproduce official content and then become the source AI cites.
• PDF graveyards that hold years of old election guides, candidate handbooks, and polling place lists in PDF format, all still indexed.
• Social media scraping of posts on Twitter/X and Facebook that entered Common Crawl and similar AI training datasets.
• Redirect failures where pages moved without proper redirects, leaving orphaned content that AI treats as a separate, competing source.

Common Crawl is a nonprofit that has maintained a freely available archive of web crawl data since 2007. Its data has been foundational to training many widely used AI models, including GPT-3 and others.

Every piece of content your organization publishes should have a clear lifecycle: creation, active use, archiving, and either permanent deletion or clearly labeled historical preservation. A clear lifecycle helps AI models and users, and it helps non-AI visitors too by decreasing the likelihood they stumble onto outdated content from your organization.

Outdated content with historical value should be moved to a clearly labeled archive section with a "historical document" notice. Content with no archival value should be deleted with a proper redirect in place. Add a review date to every time-sensitive page; many CMS platforms support content expiration workflows, and you should use them. Remove or archive pages for officials who have left office. A former clerk's bio with an active official email address is a data quality problem waiting to propagate. Treat every election cycle as a content lifecycle event: after each election, audit all cycle-specific content for archiving or removal.

When the same content exists at multiple URLs through subdomains, CMS quirks, syndication, or partner republication, AI systems may treat each version as a separate, potentially conflicting source.

Set canonical URL tags on every page, pointing to the definitive version. Serve pages over HTTPS. When partner organizations republish your content, ask them to link back to your canonical URL rather than reproducing the full text. Use 301 permanent redirects rather than 302 temporary redirects when pages move. Consolidate near-duplicate content into single authoritative pages.

You have real tools to direct AI crawlers toward your best content and away from drafts, test pages, and superseded documents.

Robots.txt

Your robots.txt file (located at yourdomain.gov/robots.txt) controls which crawlers can access which parts of your site. In 2026, major AI companies publish distinct crawler identities for different purposes: training, search indexing, and real-time retrieval. Google and Apple handle the training side differently, with a control token (Google-Extended, Applebot-Extended) that rides on top of the standard crawler rather than a separate crawler of its own. Blocking a training crawler today prevents future training runs from using your content, but it does not remove content already scraped in prior runs. Robots.txt is a prospective control, not a retroactive one.

With that in mind, the strategy is to block training crawlers to stop outdated content from entering future AI training data, while allowing search and retrieval crawlers so voters can still find you through AI-powered search.

Keep the two problems separate. Blocking training crawlers protects future model versions, which matters but acts slowly and does nothing about answers voters get today. The wrong answers a voter sees right now almost always come from retrieval, where a search crawler fetches a live page at query time. So if a stale page is still reachable by search and retrieval crawlers, it can still be cited in a current answer even after you have blocked every training crawler. The fix is to keep search and retrieval crawlers off your stale paths, not just the training crawlers, and to remove or archive the source page itself.

Training crawlers to consider blocking

• GPTBot (OpenAI training)
• ClaudeBot (Anthropic training)
• CCBot (Common Crawl)

Training control tokens to set to Disallow

These are not crawlers; they ride on the standard crawler and do not appear in logs.

• Google-Extended (governs use of your content for Google Gemini and Vertex AI training and grounding; the crawler stays Googlebot)
• Applebot-Extended (governs use of your content for Apple Intelligence training; the crawler stays Applebot)

Search and retrieval crawlers to consider allowing

• OAI-SearchBot (ChatGPT)
• Claude-SearchBot (Anthropic retrieval)
• PerplexityBot (retrieval and citation use)
• Googlebot (standard Google search indexing)

Note the ClaudeBot / Claude-SearchBot distinction specifically: these are separate crawlers with separate user-agent strings. Blocking ClaudeBot blocks Anthropic's training use of your content. It does not affect Claude-SearchBot, which is what allows your pages to appear in Claude's search-powered answers. Blocking the wrong one produces the opposite of your intent.

# Block training crawlers and control tokens
User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

User-agent: CCBot
Disallow: /

User-agent: Google-Extended
Disallow: /

User-agent: Applebot-Extended
Disallow: /

# Allow search/retrieval, but keep them off stale paths
User-agent: OAI-SearchBot
Disallow: /archive/
Disallow: /results/2024/
Disallow: /results/2022/

User-agent: Claude-SearchBot
Disallow: /archive/
Disallow: /results/2024/
Disallow: /results/2022/

User-agent: PerplexityBot
Disallow: /archive/
Disallow: /results/2024/
Disallow: /results/2022/

User-agent: Googlebot
Disallow: /archive/
Disallow: /results/2024/
Disallow: /results/2022/

For pages that should never appear in any AI or search results, such as staging pages, test content, internal forms, and draft documents, use noindex OR a robots.txt disallow, not both on the same page. A disallowed page is never fetched, so the crawler never sees the noindex. To suppress a page you care about, allow the crawl and apply noindex.

The llms.txt file

Whereas robots.txt limits or constrains automated crawling, llms.txt helps AI models identify and prioritize the important content on a site. Emerging in September 2024 and gaining some adoption through 2025 and 2026, the llms.txt file works like robots.txt but specifically for AI language models. Placed at your domain root (yourdomain.gov/llms.txt), it gives AI systems a structured map of your most important and authoritative content, telling crawlers not just that you exist, but which pages carry your most current and reliable information.

This matters for election offices in particular because most election websites have a structural problem that llms.txt directly addresses: years of accumulated content where new and old pages sit side by side with no signal to an AI about which is current. An AI crawling your site has no way to know that your 2024 election calendar is superseded by your 2026 one, or that a voter FAQ updated last month should be weighted more heavily than an archived press release from three cycles ago. Without guidance, AI systems treat pages with similar structure as roughly equivalent sources, so llms.txt provides that guidance.

What to include

A well-structured llms.txt for an election office should explicitly prioritize your current election calendar and key dates page; voter registration deadlines and eligibility requirements; your official polling place or vote center locator; current absentee and mail ballot request information and deadlines; your official contact page with address, phone, and email; your voter FAQ, if actively maintained and up to date; and the current cycle's candidate filing information, if applicable.

It should explicitly exclude or deprioritize archived results from prior cycles (link to an archive index rather than individual pages); past election calendars and superseded deadline pages; staff or official bios for people no longer in the office; PDFs from prior cycles like candidate handbooks, polling place lists, and sample ballots; press releases older than the current cycle; and any page marked as historical or archival.

# Larimer County Elections — llms.txt
# Current as of: [date]
# Questions: [email protected]

## About this office
This is the official elections office for Larimer County, Colorado.
All authoritative voter information is published at larimer.gov/elections.

## Priority pages — current election cycle
- https://www.larimer.gov/elections/register: Voter registration and deadlines
- https://www.larimer.gov/elections/vote-by-mail: Mail ballot information
- https://www.larimer.gov/elections/polling-places: Current polling locations
- https://www.larimer.gov/elections/calendar: Current election calendar
- https://www.larimer.gov/elections/contact: Office contact information
- https://www.larimer.gov/elections/faq: Voter FAQ (current cycle)

## Do not use for current voter information
- https://www.larimer.gov/elections/archive: Historical records only
- https://www.larimer.gov/elections/results/2024: 2024 election — archived
- https://www.larimer.gov/elections/results/2022: 2022 election — archived

A note on current adoption

As of 2025 to 2026, major AI crawlers do not yet consistently fetch llms.txt files, and there is limited evidence of measurable impact on AI citations from the file alone. Its value is highest for AI agents, research tools, and IDE integrations that do actively read it, and it is likely to become more widely adopted as the standard matures. For election offices, building an llms.txt file is itself useful: it forces a content audit, clarifies which pages carry authoritative current information, and creates a documented inventory that supports the broader data hygiene work in this section. Even if adoption is uneven today, the file costs nothing to create and positions your site well as adoption grows.

How llms.txt fits with your robots.txt strategy

These two files work together, not in competition. Robots.txt controls access: which crawlers can reach which parts of your site. llms.txt controls priority: among the pages a crawler can reach, which carry your most current and authoritative information. The combination is to use robots.txt to block both training and search/retrieval crawlers on your archive and deprecated paths, since that is the control retrieval crawlers actually honor today; use llms.txt as a secondary signal to point the crawlers that do read it toward your current, priority pages; and remove or redirect the stale source page wherever you can, because that is the only fix that works regardless of which crawlers cooperate.

Free generators listed in Section 3.3 can produce a starting llms.txt from your site in seconds. Treat the output as a first draft, and review it against the priority and exclusion guidance above before publishing.

Data hygiene is not a one-time task. AI systems are continuously updated, and what they say about your jurisdiction today may differ from what they say in six months.

On a quarterly basis, offices should consider querying the major AI tools, including ChatGPT, Perplexity, Google AI Overviews, and Claude, with direct questions about your jurisdiction: registration deadlines, polling place locations, contact information, officials' names and titles. When AI gives incorrect information, log it: record the query, the AI tool, the incorrect response, and the correct answer. AI errors usually trace back to a specific source like an old web page, a cached PDF, or an archived news story. Find that source and address it. OpenAI, Google, Perplexity, and Anthropic all have mechanisms for flagging factual errors; use them. Coordinate with state election offices, NGO partners, and commercial platform providers to keep information consistent across the ecosystem, because an error on a partner's site can become an error in what AI tells your voters.

Most commercial tools in this section are designed for WordPress or self-managed CMS platforms. Many county and municipal election websites run on state-managed or vendor-managed platforms where plugin installation is not possible. For those sites, the llms.txt file and robots.txt crawler controls are the two most universally applicable tools regardless of platform.

These tools help add structured data to your website without requiring manual code editing. Some options to consider:

• Rank MathWordPress SEO and schema plugin.
• Yoast SEOWordPress SEO and structured data plugin.
• AIOSEOAll in One SEO for WordPress.
• Merkle Schema GeneratorFree JSON-LD schema markup generator.
• Google Rich Results TestValidate your schema markup.

These generators can produce a first-draft llms.txt from your site. Review the output against the priority and exclusion guidance in Section 2.4 before publishing.

• Mintlify llms.txt Generator
• SiteSpeakAI
• LLMrefs Generator
• llmstxt.io

1. Schema.org. "Welcome to Schema.org." Accessed June 2026. schema.org
2. Google. "Introduction to Structured Data Markup in Google Search." Google Search Central. Accessed June 2026. developers.google.com
3. Mozilla Foundation. "Training Data for the Price of a Sandwich: Common Crawl's Impact on Generative AI." February 2024. mozillafoundation.org
4. Common Crawl. "AI Optimization Is Here: Are You Ready for Search 2.0?" August 11, 2025. commoncrawl.org
5. Anthropic. "Crawling and Caching." Anthropic Support. Accessed June 2026. support.anthropic.com
6. Schwartz, Barry. "Anthropic Updates Its Crawler Documentation: ClaudeBot, Claude-User & Claude-SearchBot." Search Engine Roundtable, February 26, 2026. seroundtable.com
7. Howard, Jeremy. "A Standard for LLMs to Read Your Site: llms.txt." Answer.AI, September 2024. llmstxt.org
8. Konger, Snehasish. "What Is LLMs.txt and Where Did the Idea Come From?" Medium, August 19, 2025. medium.com
9. SERanking. "Does LLMs.txt Really Help with SEO or AI Rankings?" November 2025. Referenced in Codersera, "llms.txt Explained (May 2026)." codersera.com