LightSite technical documentation: infrastructure, analytics and execution for AI search.

Short answer: LightSite is infrastructure and analytics for the AI-driven web, plus an execution agent. It serves AI bots a structured, queryable version of your website alongside your normal site, measures bot reads, agent actions, AI mentions and AI-referred visitors server-side, and turns those signals into approved content, structured-data fixes and outreach.

1. What LightSite Is

AI assistants are now a front door to every brand. ChatGPT, Gemini, Claude, Perplexity, Copilot, Meta AI and Google's AI experiences read websites, answer questions about them and send customers to them. Most websites were built for human eyes and search engines from ten years ago. They are heavy, script-driven and hard for a machine to understand with confidence.

LightSite fixes that from three directions at once:

Infrastructure. A parallel, machine-readable layer for your website that goes live in minutes, without rebuilding your site or changing your CMS. AI bots get clean, structured, official answers. Humans keep getting your normal site.
Measurement. Server-side, first-party measurement of what AI systems actually do with your brand: which bots read you, what they read, what they ask, which skills they call, how much they extract, what AI assistants say about you, and which human visitors AI sends you — all the way to conversion.
Execution. An AI agent that turns those signals into work: content drafts, structured-data fixes, outreach campaigns, refreshed scans and weekly plans, every one of them gated by your approval.

The rest of this document explains each part feature by feature: what you see, what happens underneath, and why it matters.

How the pieces connect

AI systems (ChatGPT, Gemini, Claude, Perplexity and more) send bots and agents to your brand.
The LightSite parallel layer serves them structured pages, skills, the catalog, Q&A and AI sitemaps, while your website keeps serving humans.
The measurement engine logs every bot request and every AI-referred visit: reads, skills, extraction, visibility, mentions, visitors and conversions.
The LightSite AI agent turns that into signals, a weekly plan, approvals and executed work.

2. The AI-Ready Website in Minutes: The Parallel Layer

2.1 The idea: infrastructure, not just rendering

Many tools "render" a page for bots: they take your HTML, execute the JavaScript and hand the bot a flattened copy. That is a copy of the same page, with the same clutter.

LightSite builds something different: a parallel layer of your website designed for machines. It looks like a sitemap from the outside — that is how bots discover it — but behind it is a live, queryable interface, much like an MCP connection for AI agents. A bot can:

read a clean structured version of any page,
ask a question and get the official answer,
look up a product, service, FAQ or proof point,
discover the full catalog of what the brand offers,
and call named skills described in a machine-readable API contract.

The result: AI systems get the facts in a format they are built to consume, in roughly half the data of a full web page.

2.2 What the layer contains

| Component | What it is | Why it matters |

| Mirrored pages | Each page on your site is read, cleaned and rebuilt as a structured page with rich JSON-LD (Organization, WebPage, Article, FAQPage, Service, Product) | Machines get unambiguous facts: who you are, what you sell, who wrote it, when |

| AI sitemaps | Dedicated sitemaps listing the structured pages, URL-encoded so Japanese, Chinese and Korean addresses work cleanly | Bots discover the layer through the standard they already follow |

| AI catalog | A machine-readable index of the brand's offerings, pages and endpoints with absolute URLs | One request tells an agent everything available |

| OpenAPI contract | A formal description of every callable endpoint | Agents and developers can integrate without guesswork |

| Skills manifest | Named capabilities an agent can call: search Q&A, look up products, fetch FAQs, read proof points | Turns the website into something an agent can use , not just read |

| Q&A search ( qa.search ) | A question-and-answer endpoint built from the site's own content | AI asks a question, gets the brand's official answer |

| Status endpoint | Plain JSON health and freshness information | Agents and LightSite itself can verify the layer is live |

All machine endpoints serve plain application/json with absolute URLs, which is exactly what browsing assistants handle best.

2.3 How pages are read: the Page Checker engine

Every page in the layer is produced by the same reader that powers LightSite's page pipeline. It extracts what a machine actually needs:

readable text after navigation, scripts and boilerplate are removed,
the heading outline of the main content (H1–H6),
title, description and canonical URL,
language (from the page, falling back to script detection),
publication date and author from meta tags, elements or the visible byline,
CJK-aware word counts (Japanese, Chinese and Korean counted by character),
whether the page already publishes its own JSON-LD,
and the resulting structured-data type and JSON-LD.

The reader also identifies pages that are unreadable to machines — blocked, timing out, too large, or JavaScript-only shells — so LightSite knows exactly which pages a bot cannot see on the original site.

2.4 Sync: always fresh

The automatic sitemap sync walks your site's sitemaps (including sitemap index files), discovers new and changed pages and rebuilds their structured versions. The Page Sync panel in the app shows what is synced, what is pending and how fresh each page is. Deterministic page IDs mean a page keeps the same address in the layer across syncs, so bots never chase moving targets.

2.5 Deploying in minutes

LightSite connects to the site you already have. No CMS migration, no rebuild, no new hosting. The app generates copy-paste-ready instructions for the platform you run:

| Platform | How it connects |

| WordPress | LightSite plugin: redirects/proxy rules, tracking and discovery aliases in one place, survives theme updates |

| Cloudflare | A Worker rule at the edge |

| Nginx / Magento | Server rules, including edge reverse-proxy setups |

| Shopify | Native integration: theme, robots.txt and routing |

| Any site | robots.txt sitemap line, discovery redirects and the tracking snippet |

The One-Click Integration screen produces the exact code for your site, and the Page Routing Coverage view shows which rules are live.

LightSite prefers 200 proxy rewrites over 301 redirects for discovery endpoints: the bot gets the content directly at your address, with no extra hop.

2.6 Channeling the bots: edge page routing

For sites that want the strongest setup, LightSite routes AI bots straight to the structured version of a page at the edge, while humans continue to the normal site:

A request arrives at your edge (Cloudflare Worker, Nginx/Magento or the WordPress plugin).
The rule checks whether it is a recognized AI bot making a public GET / HEAD request with no cookies or login, on a public path.
If so, it asks LightSite for the structured page.
LightSite returns it only with HTTP 200 and a confirmation header ( X-LightSite-Served: edge-page ).
Anything else — any other status, any delay — falls straight back to your own page.

This gives three guarantees by design:

Humans never see LightSite pages. Cookies, logins and non-bot traffic always go to your site.
Your site never breaks. Every failure path ends at your own page.
The bot list updates itself. The recognized-bot list and excluded paths are served centrally from a routing-config endpoint, so installed rules pick up new AI bots without another customer change.

2.7 Why a parallel layer beats a rebuild or a CDN swap

Easy to add. A few lines at the edge, a plugin, or robots.txt. Live the same day.
Easy to remove. Delete the rule and the site is exactly as it was. Nothing in your CMS changed.
Scoped. Turn it on for a section, a language, or the whole site.
No vendor lock-in on your stack. Your CDN, host, CMS and design system stay as they are.
Built to evolve. As new discovery standards for agents emerge, they are added to the layer centrally — your site does not need to be rebuilt each time.

2.8 Zero effect on your regular SEO

The parallel layer is built for AI bots only. It does not replace, compete with or change the website that Google and Bing rank.

Why it cannot affect your rankings

Search engines are never routed to it. Edge routing serves the layer only to recognized AI bots and assistants. Googlebot, Bingbot and every regular search crawler always receive your own pages, exactly as they do today.
Every layer page is noindex, nofollow for Google and Bing. Even if a search engine found a LightSite page, it is instructed not to index it and not to follow its links. It cannot appear in search results, cannot compete with your pages and cannot pass or dilute link equity.
No duplicate content. Because layer pages are never indexed, they cannot be treated as copies of your pages. Your canonical URLs stay the only URLs search engines see.
Your site is untouched. No change to your HTML, templates, CMS, URLs, canonicals, redirects, sitemap, robots rules for search engines or page speed for humans.
Fail-safe by design. Any error, delay or unexpected response falls back to your own page, so neither visitors nor crawlers ever get a broken page.

How it works, step by step

A request arrives at your edge.
If it comes from a regular search crawler or a human, it goes straight to your normal website. Nothing changes.
Only if it is a recognized AI bot on a public, cookie-free request does the rule ask LightSite for the structured page.
That page is marked noindex, nofollow for Google and Bing, so it stays out of search indexes.

What this means for your SEO team: your rankings, traffic, Search Console data and existing SEO work carry on exactly as before. LightSite adds a separate channel for AI assistants on top. You can verify it yourself: request any page as Googlebot and you get your own site; check the LightSite layer and you will see the noindex, nofollow directive.

3. AI Analytics: Measuring What AI Actually Does

LightSite measures AI activity server-side and first-party . Every request a bot makes to the parallel layer is logged by LightSite's own servers. Every human visit referred by an AI assistant is recorded by the tracking snippet on your site. No sampling of third-party panels, no modelled estimates.

3.1 The data foundation

| Data | What it records |

| Bot interactions | Every AI bot request: platform, bot variant, bot category, path, endpoint and skill called, HTTP status, response size in bytes, content type, timestamp |

| AI-referred visits | Every human session arriving from an AI surface: landing page, referrer, first and last AI source, number of AI touches, journey, engagement class (bounce / engaged / converted), conversions |

| Canary signals | Unique trap links planted in the layer, used to prove which agents follow through from reading to acting |

| Visibility scans | What AI assistants answer to real buyer questions about your category |

| Brand mentions | Where your brand appears across the web, with authority, link status and sentiment |

| Search Console | Your uploaded organic search data, for correlation with AI activity |

Source canonicalization

AI vendors use many domains and parameters ( chatgpt.com , chat.openai.com , Gemini grounding redirects and more). LightSite runs every source through one canonicalization table at write time and read time, so each assistant shows up as one clean bucket — ChatGPT, Gemini, Claude, Perplexity, Copilot and others — instead of splitting across a dozen raw strings.

3.2 Bot traffic dashboard

What you see:

AI bot reads over time, by engine and by day.
Engine breakdown: OpenAI, Google, Anthropic, Perplexity, Microsoft, Meta, ByteDance and others, with bot variant (training crawler, search crawler, user-triggered agent).
Most-read pages and the share of reads that hit the structured layer.
Return patterns: how often each engine comes back, and how quickly.
Bot health: share of bot requests that hit broken or redirected URLs on the original site, so you can fix dead paths AI keeps trying.

How it is measured: directly from the logged bot interactions, aggregated in the database by platform, path and day.

3.3 Extraction depth: how much each AI takes

LightSite records the exact size of every response sent to an AI bot. That turns into a metric no page-tag tool can see: extraction depth.

Average kilobytes served per request, per engine.
Tiering into light (under 10 KB), moderate (10–50 KB) and heavy (over 50 KB) extraction.
Per-site, per-day totals that reveal each engine's "personality": some take one precise answer, others bulk-read entire profiles.

For marketers, extraction depth answers a simple question: is AI taking just a headline, or is it taking the full story — pricing, proof points and FAQs?

3.4 Crawler behavior signals

Beyond volume, LightSite classifies how bots retrieve:

Bulk reads — whole-profile or catalog pulls.
Targeted calls — specific questions, product lookups, FAQ fetches.
Search queries — the actual questions bots send to the Q&A endpoint, shown in the Bot Search Queries view.

This reveals intent. When targeted calls on a topic spike, AI systems are actively trying to answer that question about you.

3.5 Skills adoption: measuring what agents use

Every endpoint in the layer carries an endpoint key and group, so LightSite can measure usage per skill :

total interactions per skill,
which platforms use each skill,
average response size per skill,
extraction-depth tier per skill.

The Skills Adoption and AI Product Requests views show which capabilities agents adopt — for example, how often qa.search is called and which questions it answers. This is the difference between "a bot visited" and "an agent used our official answer to do its job."

3.6 Agent Activity: when bots act like users

AI is shifting from crawling to acting. ChatGPT agent mode, user-triggered assistants and browser agents visit pages on behalf of a person. LightSite's Agent Activity widget records these sessions on your live site through the tracking snippet:

Agents are labelled only from a declared agent user agent or exact, value-checked automation signatures, re-checked on every action. Ordinary human behavior (autofill, keyboard shortcuts, paste) is never mislabelled as an agent.
Each session shows the agent, pages visited and actions taken.

Canary links add a second proof: when an agent follows a unique planted link, LightSite knows that specific agent moved from reading to acting.

3.7 Signed bot identity

LightSite's own verification bot signs its requests using Web Bot Auth (HTTP Message Signatures with an Ed25519 key). Its public key is published in a signature directory, so networks such as Cloudflare can cryptographically verify that a LightSite check is genuine. LightSite's site checks are therefore recognized and trusted rather than blocked as unknown traffic.

3.8 Human visitors from AI and conversions

This is where AI search meets revenue. The tracking snippet records every human visit that arrives from an AI assistant:

AI-referred visits and unique visitors, by assistant, by day and by landing page.
First and last AI source and number of AI touches in the journey.
Engagement class for each session: bounce, engaged or converted.
Conversion events and conversion rate of AI-referred traffic.
Month-over-month change, never inflated lifetime totals.

The agent's home screen shows a live status strip: AI visits, mentions, conversion rate and last scan, each with its monthly delta.

3.9 Top pages by AI performance

A page-by-page view combining bot reads by engine, extraction, AI-referred visitors and conversions, compared with the previous period. Marketers see which pages AI reads, which pages AI sends people to, and which pages turn those visitors into customers.

3.10 Health KPI cards

Four live cards summarize how AI interacts with the site: crawl coverage, structured-layer adoption, agent follow-through and sync freshness. A trust-signal widget and signals rail surface anything that needs attention.

3.11 Search Console correlation

Upload a Search Console Performance export and LightSite ingests up to 100,000 queries, in any language, with a lock that prevents overlapping uploads. The agent then correlates organic demand with AI activity:

topics with search demand but no AI visibility,
pages with impressions but weak click-through,
crawl and mention patterns next to organic clicks.

A badge in the agent shows how many days old the upload is: green while fresh, red after 30 days.

4. Measure What Matters on Every AI

4.1 AI visibility scans

LightSite asks the AI assistants your customers use the questions your customers ask, and records what they answer. Scans cover ChatGPT, Perplexity, Gemini, Claude and Google's AI Mode / AI Overviews.

Each scan records, per question and per assistant:

whether your brand is mentioned,
your position in the answer,
which competitors appear,
the sentiment of what is said about you,
the sources the answer draws on.

Query DNA and location awareness

Questions are generated from your business profile across intent layers — discovery, comparison, problem-solving and branded — and run with the country and language of your site, so a Japanese brand is tested the way Japanese customers ask.

Two scan modes

Unbiased — generic category questions, the way a new buyer asks. This is the default and shows true market position.
Contextual — brand-primed questions to test how well AI knows you.

The owner sets the preferred mode once and it applies everywhere: dashboard, agent and scheduled scans. Mixed-mode comparisons are flagged so trends are always like-for-like.

Branded ratio

Owners control what share of questions mention the brand (25% by default). This keeps mention rates honest and comparable over time.

Sentiment guard

AI sentiment labels are checked against the actual answer text. A "positive" label needs real praise in the answer; a "neutral" label is upgraded to "negative" or "mixed" when the answer contains clear criticism. Marketers get sentiment they can act on.

Reliable scoring

Mention rate is computed against the questions that were actually scored. Timeouts or service errors are excluded rather than counted as misses, and partial scans are detected and labelled.

4.2 Position intelligence

LightSite turns scans into strategy:

Share of voice against competitors across topic clusters.
Competitor landscape and trend: who is gaining, who is losing, on which topics.
True gaps — topics where you are missing or contested, not ones you already own.
Strongest wins — non-branded questions where you lead.
Page and slug authority — which of your pages carry authority on which topics, weighted by conversions, CJK-safe.

4.3 Keywords, sources and citation intelligence

Every scan records the sources AI answers draw from. LightSite classifies those domains and shows:

Where you are cited — your own pages that appear as sources.
Source influence — which third-party sites shape answers in your category: listicles, review sites, Reddit, LinkedIn, publishers, directories.
Citation network — how sources connect across topics and competitors.
Action type for each source pattern: earn a placement, publish an answer, fix a page or run outreach.

This gives marketers the map of which pages and which publications move AI answers, keyword by keyword and topic by topic.

4.4 Brand mentions and authority

LightSite scans the web for your brand weekly and on demand:

each mention with URL, authority, sentiment and whether it links to you,
claimed vs unclaimed mentions — mentions without a link are ready-made outreach targets,
new vs already-known counts per scan, with first-seen dates preserved,
competitor and exclusion controls (allowlist or blocklist) so the list stays relevant.

4.5 What consumers ignore, and what marketers should do about it

The combination of layers shows the gaps that matter:

| Signal | What it tells a marketer |

| Bots read a page heavily, but AI sends no visitors to it | The content is consumed but not compelling enough to recommend — strengthen proof and answers |

| AI sends visitors who bounce | The landing page does not match the question that brought them — rework the page |

| Search demand with no AI visibility | A topic customers want and AI does not associate with you — publish for it |

| Competitors appear in answers you are missing from | A displacement opportunity — target the sources that shape those answers |

| Mentions without links | Low-effort authority wins — outreach |

| Bots hitting dead URLs | Lost reads — fix redirects |

4.6 Why this is good for marketers

One view of every AI. Bots, answers, mentions, visitors and conversions together, across every major assistant.
First-party truth. Server-side logs and your own tracking, not panel estimates.
From awareness to revenue. Read, recommended, visited, converted — each step measured.
Comparable over time. Stable scan identity, controlled branded ratio and like-for-like modes mean trends are real trends.
Actionable. Every gap maps to a move the agent can prepare for you.

5. The LightSite AI Agent

5.1 What it is

The LightSite AI agent is an analyst and operator that lives inside the app and works on your data. You talk to it in plain language — in English, Japanese and more — and it answers from your real numbers, prepares work and executes it after your approval. Every week it also works on its own, watching for changes and preparing a plan.

5.2 How a conversation works

Understand. An intent router reads the message — including mixed-language requests, numbered choices and confirmations — and decides what is being asked: a status check, a scan, a content draft, an outreach campaign, an export.
Ground. The agent calls the right tools from a catalog of more than sixty: AI traffic, bot activity, top bot queries, visibility and position data, brand mentions, Search Console analysis, crawlability checks, page audits, content creation, outreach discovery and more.
Render real data. Numbers, queries and links come straight from tool results as pre-rendered blocks, so what you read matches the dashboard exactly.
Offer the next move. When there is a logical next step, the agent offers it. Offers are held server-side, bound to your conversation, user and workspace, and each offer maps to exactly one worker.
Execute on approval. When you say yes, the agent runs that one worker and reports the real outcome.

Before running a scan, the agent asks which assistant to scan with. Saved-data questions are answered from saved data without triggering new runs.

5.3 The autonomous loop: detectors, signals, brain, plan

data -> DETECTORS -> SIGNALS -> BRAIN -> WEEKLY PLAN -> PROPOSALS -> APPROVAL -> EXECUTORS -> OUTCOMES Detectors are pure functions that scan your data and emit signals, with no side effects:

AI traffic drop, bot silence, visibility drop, competitor surge,
mention gained on a high-authority site, negative mention, mention loss,
citation win, citation regression, new true gap,
Search Console demand without AI visibility, low click-through opportunities,
listicle and blogger outreach opportunities,
on-site audit findings: missing or invalid structured data, thin pages served to bots.

Scoring. Each signal is scored on business importance, demand, AI visibility gap, fixability and confidence. Signals confirmed by several data sources rank higher than single-source ones. Critical technical blockers (for example a page that serves bots a thin response) always rise to the top.

The brain ranks signals into a weekly plan of moves, each with a reason and a target page, and routes every move to the right kind of work:

Off-site (scans, citations, mentions, bots, Search Console): publish an article, draft a brief or run outreach.
On-site (live page audits): draft content, update structured data, or flag a site fix for your team.

Proposals. Each executable move becomes a proposal card in chat. You approve, dismiss or let it expire after seven days.

5.4 Executors: how tasks run end to end

| Executor | What it does end to end |

| Draft brief | Builds a content brief from your real scan, bot and Search Console evidence, suggesting answer formats that are proven to earn AI-referred visits |

| Publish article | Writes a full article in your brand voice for your ICP, with the strategic reasoning saved alongside it |

| Update JSON-LD | Reads the target page, picks the right schema type, builds the structured data from the page's own facts and returns a ready-to-apply draft |

| Run outreach | Discovers placements, qualifies them against your business profile, finds contacts, writes personalized emails and prepares the campaign for approval |

| Refresh scan | Re-runs a visibility scan to confirm whether a change moved the needle |

Each action type has one dedicated worker, and the agent only ever describes a move the way that worker can actually perform it.

5.5 Content creation

The content agent creates articles, briefs, FAQs, comparison pieces, social and community posts in multiple formats, grounded in your Search Console queries, scans and bot questions. It writes in your brand voice for your ideal customer, avoids generic filler, and keeps the data-driven reasoning for every piece. "Create high-impact content" turns the top gaps into ready drafts, with email delivery status reported exactly as it happened.

5.6 Outreach campaigns

Outreach puts your brand on the pages that shape AI answers:

Discover product-comparison listicles, educational guides, citation sources, bloggers and unlinked mentions in your category.
Qualify each target against your business profile; listing, tag and search pages are never targets.
Find contacts with verified email addresses.
Draft personalized, first-person emails on behalf of your company, signed with your company name.
Approve inline in chat — nothing sends without you.
Send and track replies, bounces and unsubscribes, with automatic suppression of bad addresses.

5.7 Weekly rhythm: digests, email and Slack

Monday: full weekly report — what changed, what is working, what the agent recommends.
Thursday: a changes-only pulse.
Alerts: high-priority signals the moment they appear.
Slack: the same updates in your channel, in Block Kit format.

Every autonomous email goes through a shared daily guard: one message per workspace per day, keyed by message identity, so customers are never flooded or double-sent. Opt-outs are always respected.

5.8 Learning from outcomes

The agent learns only from work that verifiably went live. An action counts once there is recorded evidence it was applied or sent — a receipt or a live page read — and only after the full outcome window from that moment. Opened or copied drafts are treated as preferences, not results. This keeps recommendations tied to what actually works.

6. LightSite in ChatGPT and Claude: The MCP Connector

Customers can connect LightSite directly to ChatGPT or Claude through LightSite's MCP connector at https://mcp.lightsite.ai/ .

16 reading tools: sites, site summary, bot traffic, AI referrals, prompt performance, recommendations, pages, brand mentions, next actions, cross-account benchmarks, page AI performance, page diagnosis, period comparison, content opportunities, authority opportunities and action status.
2 approval tools: propose an action and cancel a pending one. Proposals land in LightSite for approval; the connector itself never drafts, publishes, edits or sends.
OAuth 2.1 with PKCE (S256), dynamic client registration, resource-bound tokens, 8-hour access tokens, 60-day refresh tokens and separate read and actions:write scopes.
Enterprise control: administrators can restrict connections to their own email domain; customers see and revoke connections in Business Settings.
Freshness built in: summaries include the time of the last bot read, AI referral, scan and page sync.

Benchmarks come from anonymous aggregates with a minimum cohort of eight customers, so no names, domains or content ever cross accounts.

7. Reports and Exports

| Report | Contents |

| AI SEO report | Executive view with a performance score, AI visits, conversions, bot reads, visibility and prioritized plays |

| Monthly report | Month-in-review by email and Slack, protected against duplicates at the database level |

| Knowledge graph export | Your brand's entities, offerings and relationships in a portable file for your agency or dev team |

| CSV exports | Bot traffic, AI referrals, scans and mentions for your own analysis |

Reports are delivered in the workspace's language, including full Japanese support.

8. Built for the Enterprise

No re-platforming. Works alongside any CMS, CDN and stack. Marketing gets results without a web rebuild.
Fast to deploy, simple to remove. A rule, a plugin or a robots.txt line. Live the same day.
Multi-site and multi-language. Every workspace has its own country and language; site-aware keys isolate each property. CJK URLs, word counts and content are first-class.
Organizations and roles. Teams share workspaces under one organization; critical settings are owner-only; roles live in a dedicated table and are validated server-side.
Governed AI actions. Every executable action requires approval, expensive features are gated, daily run limits and per-workspace AI budgets keep usage predictable, and idempotency keys prevent duplicate runs.
Your assistants, your data. The MCP connector lets teams query LightSite from ChatGPT or Claude under their own permissions.
Industry-agnostic. Data-driven rules, no hardcoded verticals or keyword lists — built to work for any industry, in any language, at any scale.

9. Built for Security

9.1 Humans never touch the layer

Edge routing only serves the parallel layer to recognized AI bots on public, cookie-free, unauthenticated requests. Logged-in users, carts, accounts and private paths always go to your own site, and any failure falls back to your own page.

9.2 Data isolation

Row-level security on every table — over two hundred database policies enforce that each user sees only their own organization's data.
Tenant isolation on every query through the workspace ID, with composite unique constraints per site.
Organization-scoped aggregation so teams see their shared data and nothing else.

9.3 Privacy by design

No personal data collection: public website content and anonymous visit events.
IP addresses are hashed per workspace, so the same visitor cannot be linked across customers.
Canary tokens are protected against enumeration.
Cross-account benchmarks require a minimum cohort and return no names, domains or content.

9.4 Authentication and access

Multi-factor authentication with one-time codes, idle-session warnings and a security monitor.
OAuth 2.1 with PKCE for external assistants; atomic, single-use authorization codes; refresh tokens bound to the registered client.
Rate limiting as a single atomic database operation that fails closed; 60 tool calls per minute per MCP connection, every call audit-logged.
Server keys are held only on the server through a single switch point, enabling clean key rotation.

9.5 Verified identity for LightSite's own bot

LightSite's checks are signed with Web Bot Auth (Ed25519 HTTP Message Signatures), with a public key directory, so security networks can verify them cryptographically.

9.6 Safe automation

Executable work always requires approval.
One daily email per workspace through the shared guard.
Outreach respects opt-outs, suppresses bounced addresses and never re-contacts recently contacted sites.
Hardened functions are only ever replaced by strict supersets of their safeguards.

10. End-to-End: A Week with LightSite

Monday, 08:00. LightSite runs the weekly brand-mention scan. Detectors process last week's bot reads, AI referrals, scans and Search Console data.
Signals. A competitor surged on a comparison topic; ChatGPT sends visitors to a pricing page that bounces; two high-authority mentions have no link; a product page is missing Product schema.
Plan. The brain ranks these into a weekly plan and creates proposal cards.
Digest. The Monday report lands in email and Slack.
Approve. The marketer opens chat and approves three cards: an article on the comparison topic, a JSON-LD update for the product page, and an outreach campaign to the two unlinked mentions plus three listicles.
Execute. The executors produce the article, the structured-data draft and the campaign. The team publishes the article and applies the schema; the marketer approves the emails.
Thursday. The pulse email shows what changed.
Next weeks. Bots read the new article through the parallel layer; skills calls on the topic rise; a refreshed scan checks the comparison questions; AI-referred visits and conversions are tracked on the page. Verified, applied work feeds the agent's learning.

11. Summary

| LightSite |

| Infrastructure | A parallel, structured, queryable layer for AI — pages, catalog, skills, Q&A, AI sitemaps — live in minutes on any stack |

| Channeling | Edge routing that gives AI bots the structured page and humans the real site, with automatic fallback and a centrally updated bot list |

| Measurement | Server-side bot reads, extraction depth, skills adoption, agent activity, visibility on every major AI, mentions, AI-referred visitors and conversions |

| Execution | An AI agent that detects, plans, drafts, fixes structured data, runs outreach and refreshes scans — with your approval |

| Enterprise | Multi-site, multi-language, governed, role-based, connectable to ChatGPT and Claude |

| Security | Row-level isolation, hashed IPs, MFA, OAuth 2.1, signed bot identity, fail-safe routing |

LightSite gives AI a working connection to your official answers — and gives marketers the measurement and execution to win in AI search.

Frequently asked questions

What is LightSite?

LightSite is infrastructure and analytics for the AI-driven web, plus an AI agent that executes SEO and AEO work with your approval.

How does LightSite make a website AI-ready without a rebuild?

It builds a parallel, structured layer of your site — clean pages, JSON-LD, an AI sitemap, a catalog and callable skills — that AI bots can read while humans keep seeing your normal site.

Do human visitors ever see the LightSite layer?

No. Edge routing only serves it to recognized AI bots on public, cookie-free requests, and any failure falls back to your own page.

Does LightSite affect my regular SEO or Google rankings?

No. Google and Bing are never routed to the layer, and every layer page is noindex, nofollow for them, so it has zero effect on your rankings, URLs or existing SEO.

What does LightSite measure?

Server-side bot reads, extraction depth, skills calls, agent actions, brand mentions in sampled AI answers, AI-referred human visits and conversions.

Which AI assistants does LightSite cover?

ChatGPT, Gemini, Claude, Perplexity, Copilot, Meta AI and Google's AI experiences, each with its own crawlers.

Does the agent act without approval?

No. Every executable task — content, structured-data fixes, outreach — is proposed as a card and runs only after your team approves it.

Can I use LightSite inside ChatGPT or Claude?

Yes. The LightSite MCP connector lets your team query its AI search data from ChatGPT or Claude under their own permissions.