The ReachLLM Agent

The ReachLLM agent is the built-in assistant for brand visibility work. It behaves less like a chatbox and more like an AEO copilot: it plans multi-step work, searches the live web with sources, runs tasks in parallel when it can, and remembers reusable workflows as skills.

The Agent checks the workspace's current execution-credit balance before each response. Monthly credits and non-expiring top-up credits are combined. If you add credits or upgrade after a response was stopped, the old response remains in the conversation as history, but the warning clears as soon as the live balance can cover that response.

Authorized team members use the billing owner's execution-credit allowance. If the owner has unlimited usage, Agent responses keep recording usage without applying a credit cap. Team members still only have access to their assigned projects and existing permissions.

What it can do

The agent can:

Explain visibility, traffic, and GEO results.
Answer questions about your own account: which plan you are on, how many execution credits you have used and have left (including non-expiring top-up credits), your tracked prompt, GEO audit page, and project limits, and which AI models are tracked for you. These come from your live billing record, so they match what Account > Billing shows.
Read connected AI crawler traffic, including crawler totals, top bots, top pages, daily trends, and recent visits for the active brand.
Explain saved Bot View page text, access rules by purpose, and browser differences, or run a bounded free inspection when you ask.
Compare two visibility runs and tell you exactly which prompts gained or lost a mention for a specific AI platform (for example, "which prompts dropped for Perplexity between June 24 and the latest run"), and separate a real lost mention from a platform that simply was not queried that run.
Turn Brand Visibility answers into inline charts for visibility score, share of voice, average rank, mentions, citation rate, trends, and competitor comparisons. Use the metric buttons to change the chart and View data to inspect the values in a table.
Handle a rebrand. Ask it to change your brand name, website domain, or alternate names and it updates the existing brand in place, so you do not end up with a duplicate. Keep the old name in your aliases: AI platforms trained before the rebrand will still use it, and the aliases are what keep those mentions credited to you. GEO audits and visibility runs from before the change still refer to the old domain, so the next run is your first clean baseline on the new one.
Run or inspect audits. For new page audits, the Agent reports the deterministic core issue score, passed, failed, and not-counted totals separately from supplementary GEO checks. It keeps historical overall health clearly labeled and does not treat a not-counted check as an issue.
Platform admins can update Brand Visibility model preferences in Account settings or ask the Agent to switch models directly. ChatGPT choices use GPT-5.6 Sol Medium, High, Extra High, or Pro. Gemini choices use Gemini 3.7 Flash, Gemini 3.6 Flash, or Gemini 3.1 Pro. Retired model names are not accepted.
Generate prompts, content, schema, and llms.txt.
Search the live web for current information and cite its sources.
Read connected context from supported integrations.
Read existing website content and publish or prepare changes through WordPress, GitHub, Webflow, or Shopify when connected.
Preserve platform-specific edit details from connected CMS reads, including GitHub paths and SHAs, Shopify asset IDs, and Webflow editable node IDs for targeted component updates.

Find existing citation opportunities

Ask the agent to "Find non-branded questions where our brand is already mentioned or cited." It can quote a discovery run, prepare up to 25 natural conversational prompts, check every exact question across all the brand owner's enabled AI platforms in the background, inspect matching answers and sources, and add a specific suggestion you choose to tracking. The same feature is available in Your Prompts > Discover prompts for every brand.

The agent can switch background discovery on daily or weekly, or turn it off. Before a manual run or schedule activation, it shows the current platforms and maximum credit budget for your approval. It must use that exact quote and cannot silently accept a higher cost or changed platform list. Background discovery saves live matches for review and never promotes them automatically. These actions use your current brand permissions, execution credits, and tracked-prompt allowance, including pooled workspace limits. Team members with view-only access can inspect saved evidence.

Recorded citations, saved page topics, brand context, and connected crawler page activity can supply topic seeds. Discovery creates a diverse set of natural conversational questions from those seeds, preserving recorded wording separately from generated wording. Only supported buyer context is used; generated phrasing does not establish measured AI conversation demand. They do not count as current visibility. Questions containing saved own brand or distinct product names are excluded, including spaced spellings and model numbers. Generic offering labels remain valid discovery topics. Third-party app and media catalog listings do not count as owned website citations. The agent reads saved discoveries that originally passed a real answer check for a brand mention or owned website citation. Genuine saved positive evidence stays available across later negative or incomplete checks, age and changed platform settings. It must report the original positive date and distinguish an earlier check from current visibility. The latest scan status is separate from the last positive evidence. It distinguishes a written mention from a citation, preserves each platform's checked date, and can retrieve one chosen platform's full saved answer. The in-app evidence preview focuses on the matching passage, with the full answer available to expand. Unverified historical candidates and unanswered checks do not become positive findings. Previously verified positive answers remain saved with their original evidence and date. Source links are preserved in the saved answer, with readable page addresses when source titles are unavailable. A new run can recheck questions whose earlier citation evidence was incomplete. API results can differ from individual consumer chat sessions, and discovery cannot recover private user prompts.

The agent can archive a specific saved discovery when you explicitly ask to hide it, and restore it when you ask to bring it back. It must not archive findings automatically after a new scan or a negative answer. Archived discoveries retain their evidence and stay out of the main review list and future discovery checks until restored. Archiving and restoring do not spend execution credits, change the schedule or archive a tracked prompt. The agent can read saved or archived discoveries in pages and retrieve one chosen finding directly. These actions use the same brand permissions as the app. New-run logos and checks use only the owner's currently enabled platforms; historical evidence can retain a formerly enabled platform as an earlier check.

In Your Prompts > Discover prompts, View last scan shows the exact tested questions and completed or incomplete checks, including runs with no matches. These checks are not proof of prompt demand. Discovery also rejects obsolete system-file questions when reusing saved candidates. The agent uses the same candidate validation as manual and scheduled discovery. Run citation discovery starts a durable background job and returns its queued status promptly. The agent can read lightweight progress through Get citation discoveries with statusOnly: true, then fetch findings when checks finish. Closing the app does not cancel the run. A queued or running result does not establish that no matches exist. The agent must read progress rather than starting another paid scan. The exact quote supplies the maximum question count, cost and enabled platforms.

A run uses the same 20 execution credits per prompt as a manual Brand Visibility refresh, covering all enabled platforms together. Research and question preparation have no additional credit charge. The 25-prompt maximum is 500 credits. The maximum is reserved once and settled after all background checks finish. A prompt uses credits when at least one platform returns a completed answer, including a negative answer. Unused reserved credits are returned, including prompts where all platforms failed. Preparation-only runs and runs with no completed platform answers cost zero execution credits. Historical receipts preserve their original amounts.

Manual and scheduled discovery use the same budget policy. Removing the research fee preserves matching existing 25-prompt schedules at the lower cost and keeps their next run dates. Changes to brand scope, platforms or a higher budget still need review. The agent must obtain a fresh quote for each new run instead of hardcoding a cost. The same per-brand monthly scan limit and cooldown apply to the UI, schedules, and agent. Adding an explicitly chosen suggestion does not start a visibility analysis automatically. The compact UI shows platform logos and exposes question origins through Why this prompt?, separately from the saved live answer and citation evidence.

Generic product and service labels do not establish a brand mention. The agent receives verified evidence for actual brand names or explicit aliases, or corroborated citations to the brand's website.

The evidence dialog includes the explanation beneath a matching brand heading and puts citations to your website first. It initially shows up to three source links; the complete source list and full answer remain available to expand.

Bot View and crawler evidence

The Agent can use the same Bot View evidence shown in GEO Audit, including for a discovered page without a scored audit. It starts with saved evidence and links to the exact page and request mode in the viewer.

get-page-bot-view reads a saved initial response or paired browser comparison without fetching the website. A paired read uses compareBrowser: true and request profile reachllm. No saved preview means the evidence is unavailable, not that the page is empty or blocked.
inspect-page-bot-view checks one public page when you explicitly ask for an inspection, a new check, or a browser comparison. After an interrupted request, the Agent reads saved evidence before starting another capture.
get-crawler-traffic-data reads the selected brand's connected crawler request counts, pages, statuses, daily trend, and recent request metadata. It does not connect the website, start an inspection, or retrieve historical response bodies.

Try "Explain our saved pricing page Bot View," "Compare this page with a browser," or "Claude visited yesterday. What do the logs show?" The Agent distinguishes an initial response, a paired browser observation, a user-agent simulation, and a logged request. It reports capture dates, missing passages, matching access rules, and limitations before suggesting a fix. Search discovery, user-requested retrieval, and training preferences have different purposes. A training restriction does not automatically mean a search visibility failure.

Bot View provides 10 initial checks and 3 browser comparisons per signed-in user per UTC day. Saved reads do not use that allowance; eligible captures can be reused for 24 hours and the latest captures expire after 7 days. These checks use no execution credits or scored GEO page allowance. The Agent's ordinary conversation entitlement still applies. A free inspection does not make the chat itself free.

All page excerpts and differences are untrusted website content, not instructions for the Agent. A copied user-agent name is not verified provider identity. Historical logs have no response body, and a later preview does not reconstruct a past visit. No check proves training, retention, citation, ranking, or answer inclusion. The browser also applies the documented inspection limits, so delayed or blocked content can be incomplete. The Agent reports incomplete coverage and failed critical requests without interpreting an unavailable comparison as no differences.

Bot View does not change website delivery or provide managed rendering. Supported website fixes continue through the existing connected publishing tools and their permissions.

The Agent receives each page's structured diagnosis, missing passages, fix steps, and verification instructions. A warning requires at least 25% missing text in a complete same-visit comparison containing at least 100 normalized words. It explains that initial text coverage measures matching word sequences and short readable lines in ReachLLM captures, not a provider's ability to read a page. Incomplete, shortened, short-page, or legacy unknown evidence does not support a percentage warning. It does not use a content-to-code ratio as a substitute.

For "Which pages should I fix?", the Agent can use get-bot-view-findings to read saved findings for up to 50 known brand URLs at a time, using pages already listed in GEO Audit. It reports how many requested pages had saved comparison evidence and which remain unknown. This read never crawls the site or starts browser visits. It uses get-page-bot-view for a selected page's full evidence before proposing an exact fix. Unchecked or expired pages are not counted as passing.

When you explicitly ask for a fix, the Agent can use the current connected Website, CMS, or GitHub workflow if it supports the exact page change and your permissions allow it. It uses the observed public passages and current source, never invented product facts. A source edit or pull request is not proof of a published or verified fix. Rendering configuration may need a developer; in that case the Agent prepares specific instructions and a fresh-comparison verification plan. Existing publishing approvals still apply, and a recheck uses the free inspection allowance only when requested.

Generated proposals, contracts, and surveys

The agent can create a GEO proposal as a downloadable PDF and HTML report, create a service agreement in Google Docs, and create a live onboarding survey when the required integration is connected.

Follow-up feedback in the same conversation updates the saved deliverable instead of starting over:

Proposal edits reuse the active proposal and regenerate its PDF and HTML downloads. The edit is scoped to that proposal only, so it does not change the master template, another customer's report, or a future proposal.
Contract edits update the same Google Doc and keep its existing link.
Survey edits update the same live form and keep its existing share link.

You can ask for another change in a later message because the conversation remembers which proposal, contract, or survey is active. Proposal changes also create version history, so you can ask the agent to list or restore an earlier version.

Working in the chat

Attach files (CSV, PDF, text, images) with the plus button. Attached files show as chips on your message; click a chip to preview the content or download the original file. Images are included in the current multimodal Agent turn, while documents are retrieved through the file pipeline.

Working with the agent inside Content Studio

Content Studio uses the same ReachLLM Agent with structured document context. Select a passage and press Command-I (or Control-I) to open the side panel with the selected block, text range, surrounding sections, document ID, and current revision attached. The selection appears as a quoted attachment in the composer and remains visible on the sent message.

The agent can read the draft alongside your approved brand facts, writing style, competitor set, website content, connected search data, and AI visibility research. When a request changes the document, the agent creates a reviewable proposal instead of replacing the article. Accept applies the proposal to the current shared document and records a version. Reject keeps the document unchanged. If the selected text or revision changed while the agent worked, ReachLLM marks the proposal stale and asks the agent to retry against the latest draft.

The embedded Agent uses the same files, voice input, competitor mentions, saved skills, slash commands, Plan mode, connected tools, thinking activity, sources, charts, and follow-up actions as the main Agent page. Each document has one shared Agent conversation, so teammates can watch and continue the same live session while unrelated chats remain separate.

Running a saved skill uses the same composer path as a typed request. The skill keeps the active Plan setting and selected tools, then uses the current document and selected passage as context.

When Plan mode needs approval, the Agent explains the plan first and shows Proceed with the plan or Adjust something below that explanation. Approving a plan does not bypass the document proposal and review controls.

Content Studio also gives the agent dedicated tools to:

list and read Content Studio documents
create a reviewable suggestion against a specific revision
start basic or deep review
read and resolve review findings
find internal links from real website, audit, and Content Studio page context
request a brand-aligned image
move an approved revision through publishing
send an approved, revision-pinned website document to the Website workspace

Comment threads show profile photos, replies, mentions, and who resolved the discussion. The release panel also shows who approved and published the current revision. Review, link, research, and asset tasks can run while teammates continue editing because they create durable findings and proposals instead of locking the document.

Paste long text freely. Anything over a screenful becomes a "Pasted text" attachment chip instead of flooding the composer, and long pasted messages stay collapsed in the thread with a Show more toggle.
Type / to open the command menu: run a web search, turn on Plan mode, or run one of your saved skills.
Type @ to mention your brand or a competitor. This works on the first screen and mid-conversation, in both the Agent page and the floating assistant.
Every answer ends with follow-up suggestions you can click to continue the thread.
The chat follows new text while you are at the bottom. If you scroll up to read, it stops jumping and shows a button to return to the latest.
The same chat is available on the Agent page and in the floating assistant on every other page.

Conversation history

Conversation history is saved separately for each brand. Switching brands shows that brand's chats, and returning to a brand restores its most recent conversation. Scheduled agent runs also create conversations in the same history so you can open the prompt and result later.

If history cannot be reached, the sidebar shows a connection message instead of incorrectly saying that no conversations exist. Refresh the app to retry the connection.

Live progress

While the agent works, it shows plain-English steps such as "Reviewing your brand knowledge," "Comparing visibility reports," or "Preparing your visibility chart" instead of a generic thinking label or raw tool names. Completed steps get a check and read in the past tense; the current step shows the ReachLLM mark. After the answer is ready, the activity summary explains what was completed and how long it took. Click the summary to review the recorded activity details.

When you ask whether a Visibility run or GEO audit finished, the agent checks the saved run and its results. Queued, running, failed, and incomplete work stays distinct from completed work. Missing results are not a zero score, and an older completed report does not prove a newer request succeeded. An audit started from the agent checks the homepage and website-wide settings. Its reply identifies that one-page scope and the saved run; discovering other pages and auditing them are separate actions.

GEO summaries include the saved audited pages, their technical and content scores, and their issues. Passing and unavailable checks do not count as problems. The core issue score remains separate from supplementary GEO health. Visibility summaries count the AI answers that mention your brand once per answer, even when an answer repeats the brand name, and show the available platform breakdown.

Inline Brand Visibility charts

For a shared brand, Brand Visibility runs requested through the agent use the brand owner's enabled AI platforms and model choices. A teammate's personal settings do not change that coverage. The shared Brand Visibility page shows the same providers and results to everyone with access when the same dates and filters are selected.

Ask for a trend, a competitor comparison, or a filtered view of your Brand Visibility data. When chart-ready data is available, the answer includes a ReachLLM visualization directly below the explanation. The same chart works on the full Agent page, in the floating assistant, and in a shared chat.

The chart uses the stored Brand Visibility reports behind the answer. Average rank reflects the order in which the selected brands appear in each answer. Missing rank or citation data stays unavailable instead of appearing as zero, and historical values need saved evidence. The chart does not estimate missing values. Filters such as lookback period, prompt selection, provider, or prompt type stay attached to the visualization, and the Open link takes you to the full Brand Visibility workspace for deeper analysis.

Competitor comparisons and agent-created visibility reports use the competitors marked Direct in Knowledge Base by default. Share of Voice, rank, comparison tables, and report summaries use that same group so the values match the Brand Visibility workspace. Choose All tracked in Brand Visibility when you deliberately want the broader discovered-company view. Sentiment does not use the page comparison switch. Its Head-to-Head card always compares the brand with one Direct competitor.

In Sources & Opportunities, open Gap Analysis and choose Plan a fix on a missed prompt to send its competitors, AI platforms, and evidence to the Agent. The Agent reviews the gap and proposes the highest-impact content or authority work before making changes.

When the agent explains or reports a Brand Visibility metric, it should include the metric name, exact value, report date or date range, active filters, comparison group, denominator, and calculation basis. A change between two rates should be labeled in percentage points. The agent should not compare competitor sentiment with the brand's sentiment because those values are not produced by the same evaluation method.

For sentiment and brand perception, the agent includes the saved analysis date, platform coverage, and evidence quality. When you ask for a current sentiment check, the agent reads the latest saved analysis before answering. It identifies unavailable data and treats earlier chat replies as history. Every saved answer is checked, and attributable brand answers receive a complete-answer sentiment assessment. The agent distinguishes the number checked from the number scored: answers without the brand are excluded, not assigned a neutral rating. Comparative context can inform an answer assessment while each business keeps its own claims and evidence. Unusable responses, uncertain business identity and unresolved assessments remain explicit. Recovery revisits the evidence behind a rejected assessment and preserves successful assessments. The agent can request the current saved-answer assessment when an older sampled result is still present.

Strengths and concerns retain readable quote text and their supporting evidence. Partial evidence is labeled explicitly, and older sentiment methods are identified before comparisons. The score averages assessed brand answers, with mixed or neutral framing contributing 50, positive framing 100, and negative framing 0. Missing answers, answers without the brand and uncertain business identity affect coverage, not the numeric score. Score explanations must refer to the assessed answers; a later review of saved answers does not make the original visibility data newer. Any explicit overall provider rating, such as 7/10, is reported separately. The agent must not describe a positive score as proof of perfect service quality, an absence of concerns, or guaranteed success on future provider requests. Topics reflect the actual business and evidence; no fixed number of topics is promised. New reports preserve the assessment coverage and evidence available when they are created. They distinguish the original analysis date, sentiment review date and report creation date.

Sentiment percentages describe assessed full answers, not the number of positive or negative statements inside them. A rounded 0% negative share does not prove that no individual criticism exists. Agent-generated sentiment tables preserve the saved topics, percentages and sentiment labels; combining neutral and negative topics must not turn their total into a negative percentage. The agent checks these distinctions and the original and review dates before showing its finished sentiment report.

Short requests such as "latest sentiment" also read current saved data. Asking for a read without a new run does not start paid visibility checks or regenerate sentiment. The result belongs to the selected brand; if you explicitly ask about a different brand, select that brand before checking it. Missing scheduled platform or prompt answers stay separate from recorded answers, and an older run with no original schedule cannot establish that all intended checks ran. A failed read is reported as unverified rather than replaced with an earlier chat score.

Current sentiment reads follow the same completed full analysis and prompt inventory used by Brand Visibility. In-progress and exploratory runs cannot replace that report. When a provider retry changes saved answers, the agent must wait for their updated review or say that the current assessment is unavailable. Previous scores and claims cannot stand in for changed evidence. An unresolved assessment has a bounded second review with the same evidence safeguards. Missing citations or uncertain business identity must remain explicit after that review.

Shared category lists and technology relationships can support an answer assessment while staying separate from company-specific strengths or claims. Being listed in a category is neutral context, not an endorsement. Joining a team does not transfer ownership of earlier brands: sentiment reads use the selected brand's own saved analysis, and shared access still requires current permission.

Brands in each of your authorized workspaces keep their own saved sentiment and reports. Switching workspaces does not transfer brand ownership or conversation history. The agent checks your current permission for the selected brand before reading or creating a report.

Try requests such as:

"Show my share of voice over the last 90 days."
"Compare our visibility with our top three competitors."
"Chart our citation rate for these product prompts."
"How did mentions change between the previous report and the latest one?"

Web search and sources

When you ask about something current (recent news, prices, competitor or industry updates), or when you share a link, the agent searches the live web and grounds its answer in what it finds. The sources appear as clickable chips under the answer, each with the site's icon and a link you can open. You can also force a search from the / menu with /websearch.

Plan mode

For multi-step work the agent lays out a short plan first, shown as a checklist that ticks off as it goes. It decides on its own when a request needs a plan. You can also turn on Plan mode explicitly (the Plan button, or /plan) to have it propose a plan and wait for your approval before doing anything.

For large requests whose parts are independent (for example: audit the site, pull visibility, and compare competitors), the agent runs those parts in parallel so the whole thing finishes faster.

Skills

A skill is a saved, reusable workflow. Ask the agent to "save this as a skill" (give it a name), or save one after it builds a workflow you like. Your skills then appear in the / menu, and you can run one anytime by picking it there or asking for it by name. The agent runs the saved steps and reports the result. Skills are separate from memory: memory is facts the agent recalls about your brand; skills are workflows it executes on demand. You can also ask it to list or delete your skills.

Speed

Quick messages like a greeting get a fast, lightweight reply. Real tasks get the full pipeline with your brand context and reasoning. You do not choose a model; the agent picks the right one automatically.

Grounding

The agent grounds every answer in your full brand knowledge base (brand details, competitors, structured profile, saved facts, and E-E-A-T signals) and your connected integrations when available.

Best uses

Ask the agent for a concrete outcome, such as:

"What should we fix first?"
"Generate schema for this page."
"Create an article for this prompt."
"Search the web for the latest on Generative Engine Optimization and cite sources."
"Save this as a skill called Weekly Visibility Check."
"Open a GitHub PR for this llms.txt update."
"Read my Shopify homepage content before rewriting it."