Every workspace in this city gets read twice: once by the person using it now, and once by whoever inherits it on a date that is already on a calendar. The second reader is the one worth writing for.
Most advice about running websites as a team treats the team as the constant and the work as the variable. Here it is reversed. Properties outlive staff, staff rotate on published schedules, and what carries a program across a handover is whatever got written down along the way.
Three departures you can already put in a calendar
Elsewhere, turnover arrives as a surprise: a resignation letter, four weeks of notice, a scramble. Here the significant departures are announced years ahead, in three flavors that have nothing to do with anyone quitting.
An administration changes and the readers your public affairs pages were written for are replaced wholesale, vocabulary included. A board rotates and the officers who approved the messaging hand their seats to successors with different priorities. A contract goes to recompete and the agency of record — holding every login and every rationale — either wins again or hands the lot to a competitor on a stated Friday.
- The date is known before the work starts. A three-year contract with two option years has an end you can name the day it is signed.
- The successor arrives without context and with authority. They can reverse a year of decisions in an afternoon, and will, if nothing explains why those decisions exist.
- Nobody schedules an overlap. The outgoing team is usually gone before the incoming one is briefed, so the file is the briefing.
- Repeating discarded work is the standard failure. The abandonment was a conversation, and conversations do not survive a handover.
The project feed as a handover document being written continuously
The Stream area is one chronological feed per project. Into it go the assistant's answers, generated reports, each new backlink with the donor's authority score and traffic figure, open to-dos and campaign news — interleaved in the order events occurred rather than sorted into four systems.
That interleaving is the feature, easy as it is to mistake for untidiness. A placement three entries below the question somebody asked about that domain, itself below the report that prompted the question, is a causal chain a successor can read. The same three items in three tidy systems are orphans.
The project feed
One record per project, holding the events and the reasoning around them together.
- Placements arrive with their evidence attached. Each link carries the donor's authority rating and traffic, so a successor can judge it without re-researching it.
- Reports land in the timeline, not in an inbox. A generated report sits at the date it describes, beside whatever was happening that week.
- Batches are accepted whole. Keyword and address lists can be handed over in bulk rather than typed in one at a time.
None of this asks for discipline from the operator, which is why it works. Documentation that depends on somebody choosing to document loses to a busy quarter every time, while keeping the campaign and its record in one place produces a file that exists on the day it is needed.
Zero to three blocks: an answer that can be traced back
The assistant here is not a general chatbot with your domain named in the prompt. It is wired to the project's data, and a router model decides per question which blocks to load — search console figures, results-page tracking, campaign state, custom material — from none to three, by relevance to what was asked.
For a skeptical reader that range is the most interesting thing in the product. Zero blocks means the question was about method and the answer is generic, which is an honest thing for an answer to be. Three means the model read the record before speaking. That is the difference between a citation and an improvisation, and an assistant tied to the project's own figures is the version worth quoting to a client.
| What somebody asks | Roughly what gets loaded | How far the answer can be trusted |
|---|---|---|
| “What is a canonical tag for?” | Nothing from the project | General guidance, no claim about you |
| “Which of our terms slipped last month?” | Logged search data, rank tracking | Reproducible from the same records |
| “Why did placements pause in May?” | Campaign state | Reads the campaign, not the analytics |
| “Are we losing to the same three firms?” | Rank tracking, campaign, custom | As current as the last sync |
| “Should we contest this term?” | Whatever is relevant — and still not enough | A prompt for a decision, not the decision |
Answers stream back token by token, and the conversation keeps up to twenty messages of history. Streaming lets a long answer be read and interrupted rather than waited on. The twenty-message ceiling makes the conversation working memory, not the archive — the durable record is the feed it lands in.
All, Links, Files, To-do — and the state that carries the most information
A feed that accumulates everything needs a way to be read narrowly, so four filters sit across it. All is the chronological record; Links isolates placements; Files isolates documents and exports; To-do isolates the work queue, and the queue is where the handover value concentrates, because each item carries a state.
All
Everything in sequence, and the only view that shows cause next to effect.
- Answers, reports, placements, tasks, news
- Read as a narrative, not scanned
Links
Placements only, each with the donor's authority and traffic figure.
- Reviewable without opening the campaign
- Outlasts whoever commissioned it
Files
Documents and exports — where last quarter's exhibit is found again.
- The trail behind a circulated number
- Reopened more often than created
To-do
The work queue, each item active, deferred or dismissed.
- Open work, visible without a meeting
- Closed work, visible with its reason
The three states repay attention. Active is the obvious one and the least interesting. Deferred says a team looked at something, judged it real and chose a later moment — exactly the item a successor would otherwise rediscover and treat as new. Dismissed says somebody considered it and declined, and that is the most expensive knowledge to lose, because nothing in a fresh workspace suggests the idea was already tried.
Full-text search runs across every message, which turns the record from something you read into something you interrogate. The question a new account manager asks in week one is never “show me the timeline.” It is “did anyone ever look at the licensing pages” — a search, answerable in seconds against a year nobody was present for.
Share by site, revoke by date
The workspace is multi-tenant in a specific way. Google accounts link into groups, so one login reaches several verified properties without juggling credentials. Individual sites are then shared to particular email addresses instead of the whole account changing hands, and that sharing can be withdrawn.
The conventional practice is to review access when somebody leaves. Where the agency of record changes on a scheduled recompete, that instinct is wrong and reliably a quarter too slow. Anchor it to the contract instead: access begins at the start date, is reviewed at each option year, and ends on the end date, whether or not anyone has gone anywhere.
| Event with a known date | What should happen to access | What usually happens instead |
|---|---|---|
| Contract award, day one | Sites shared individually to named addresses | The whole account password is emailed |
| Option year exercised | Access confirmed and the list re-read | Nothing; the list keeps growing |
| Subcontractor rolls off | That address revoked, feed left intact | Access forgotten for two years |
| Recompete lost | Access ends, record stays with the client | Both leave with the departing firm |
| Board or officer rotation | The new approver is added, not substituted | The old approver's account is reused |
Two properties matter more than the mechanics. Sharing a site and sharing an account are different acts, and only the first can be partly undone, which is why granting access one property at a time earns the extra minute at setup. And revocation removes a person's reach, not the project's history: the feed belongs to the workspace, so their reasoning outlasts their access to it.
Why flat labels beat a folder tree
Site tags act as a global filter: apply one and every screen narrows to the properties carrying it. The temptation, for anyone who has run a shared drive, is to build a hierarchy — client, practice area, jurisdiction, year. Resist it, for the same succession reason in another costume.
A deep tree encodes an organization chart, and organization charts are precisely what changes at a handover. Rename the practice areas and every path is wrong, leaving a structure that describes a firm which no longer exists. Flat tags carry no such assumption. They intersect instead of nesting, so a property can be tagged for its client, its jurisdiction and its status at once, and any one can be retired without disturbing the others.
- Tag by what the site is, not who manages it. Ownership changes on schedule; what a property does for the organization mostly does not.
- Keep a status tag current. Active, dormant, in review — the label that stops a successor spending a week on a site nobody intends to keep.
- Give jurisdiction its own tag. District, Virginia and Maryland pages answer to different rules, and here that filter earns its keep.
- Keep the vocabulary short. A dozen tags everybody uses beats forty invented independently and never reconciled.
The report is an exhibit, so it has to name its own scope
A report produced here rarely stays here. It goes into a board packet, a renewal memo, or a recompete response as evidence the incumbent did the work. Once attached to something, it stops being a chart and becomes an exhibit, and exhibits are held to a different standard than dashboards.
An exhibit states three things before anyone reads a number: which properties it covers, what date range it spans, and which record produced it. A headline with no scope is not persuasive to this audience but a liability, because the first question in the room will be one the document cannot answer.
The report builder
For figures read outside the room they were generated in, months after the fact.
- Two formats for two purposes. Machine-readable exports for the analyst re-deriving your figure; a rendered document for the packet.
- Rendering happens server-side. The file in the appendix looks the same to everyone who opens it, which matters when it circulates.
- Branding is a provenance mark. Logo and colors matter less for presentation than for keeping a file identifiable a year later.
- The row ceilings decide the format. Anything past 250 rows is a data file with a summary attached, not a document.
The on-screen visualizations serve the working half of this. Time series show direction, metric cards carry headline figures, tables filter and sort at fifty to two hundred rows a page, sparklines compress a trend beside a label, and heatmaps break results out by country and device. Each is a way of finding the finding; none is the exhibit. The export a builder produced is the exhibit, because an export can be reopened and checked.
A week in the workspace, and where the automation stops
None of this needs a daily ritual. Background workers keep syncing while nobody watches, so dashboards are current whenever they are opened. What a team supplies is judgment at a handful of moments.
| When | What the workspace has done | What a person has to do |
|---|---|---|
| Monday | Weekend movement is synced and flagged | Scan the feed, filter to To-do, triage |
| Midweek | Candidate terms surfaced, placements logged | Approve, reject or defer — with a reason |
| Thursday | Data assembled for whatever is being asked | Ask the assistant, keep what is worth keeping |
| Month end | Report generated to the configured shape | Check its scope, then send it as an exhibit |
| Contract dates | Nothing — it does not know your dates | Re-read the sharing list, revoke what expired |
That last row is the honest one, and it generalizes. The workspace automates collection, arrangement and recall. It does not automate the four things that decide whether a program is any good, and pretending otherwise survives no meeting in this city.
Prioritization
A ranked list of opportunities is not an ordered list of what to do this quarter.
- Rests on commitments no tool sees
- Moves with the leadership that set them
Finding the cause
The record shows a decline with a date. Whether a redesign, a regulation or a rival caused it is inference.
- Needs context from outside the panel
- Wrong causes produce confident wrong fixes
Setting the goal
Nothing in the data proposes a target. A target is a commitment somebody is later measured against.
- Defensible before it is set
- In writing, so it can be inherited
Talking to the client
Explaining a flat quarter to a board, in their language, with the limits stated.
- Judgment about what to volunteer
- The part that earns the recompete
Read the division that way and the tooling stops competing with the practitioner. It removes the reconstruction — two days spent working out what was done last spring, and by whom — and leaves those hours for the four cards above.
Questions that come up when the workspace is set up
We are the incumbent agency. Why would we build a record that makes us replaceable?
Because the recompete is decided on evidence, and a firm producing a dated, sourced record of three years of decisions argues from a stronger position than one offering a narrative. Withholding it protects nothing: the client owns the properties, and an agency seen to hold institutional memory hostage rarely survives the review that notices.
How do we hand a workspace over without handing over our other clients?
Share sites individually rather than granting account-level access — the only arrangement that can be partly undone. At handover you revoke the departing addresses and add the incoming ones; the feed, the to-do states and the exports stay attached to the properties.
Can the assistant answer a question about something that happened before we arrived?
It answers from the project data available to it, but for “what did we decide” questions the reliable instrument is full-text search, which returns the entry itself with its date. Use the assistant to interpret and the feed to establish what was said.
Our board wants a single quarterly document. PDF or the data export?
Both, labeled differently. The rendered document goes in the packet at 250 rows or fewer, with its scope and window on the face of it. The CSV or JSON, up to 10,000 rows, goes to whoever checks the arithmetic — in this town somebody will — and offering it unprompted removes an objection before it is raised.
How many tags should a portfolio of a dozen sites carry?
Fewer than you first want. A working set is usually client or program, jurisdiction and status — three axes, applied flat, intersecting when filtered. Resist encoding internal structure: that structure is on a rotation schedule and the tags should outlast it.
The starting point is not a campaign but an inventory. Connect the properties, tag each one, list who holds access, and write the contract end dates beside those names. Most organizations find two things in that hour: a site nobody will claim, and access held by an address that stopped being relevant several renewals ago. Open a workspace and connect the first domain, then read the rest of the blog for the analytics side of the argument. The date you are writing for is already on the calendar, and it is not moving.